Thesis Heidi Silvennoinen

Research Collection
Master Thesis
Non-Spatial and Spatial Statistics for Analysing Human

Perception of the Built Environment
Author(s):
Silvennoinen, Heidi
Publication Date:
2018-09
Permanent Link:
https://doi.org/10.3929/ethz-b-000290292
Rights / License:
In Copyright - Non-Commercial Use Permitted
This page was generated automatically upon download from the ETH Zurich Research Collection. For more
information please consult the Terms of use.
ETH Library
Non-Spatial and Spatial Statistics for
Analysing Human Perception of the
Built Environment
Thesis Elective
Submitted by:
Heidi Silvennoinen
BSc Architecture, Aalto University
Submitted to:
Chair of Information Architecture
ETH Zürich
Advisor: Dr. Varun Ojha
September 2018
ACKNOWLEDGMENTS
I am thankful to Prof. Dr. Gerhard Schmitt, Professor of Information Architecture for the opportunity
to write this thesis. I would also like to thank my supervisor, Dr. Varun Kumar Ojha, for his valuable
advice regarding the thesis. Thanks to his help and support, I have learned a great deal about data
analysis techniques and research work during this project.
I am very grateful to Dr. Saskia Kuliga, Martin Bielik and René Weiser of Bauhaus-Universität
Weimar, Germany, as well as Dr. Varun Ojha and Danielle Griego of ETH Zürich, for their work
on the Virtual Reality experiment’s design and data collection. I am also grateful for the ESUM
project at the ETH Chair of Information Architecture. Without the ESUM project and the work of the
aforementioned researchers, this research would not have been possible.
2
ABSTRACT
The world is rapidly urbanising, which makes the liveability of cities an important issue in our times.
New sources of data on urban citizens including social media, video footage, and even fitness data,
are also becoming increasingly available. This study is motivated by these two trends. We analyse
data gathered in a Virtual Reality experiment to find correlations between features of the urban
environment and people’s perceptions of this environment. In particular, the aim is to further our
knowledge of which kinds of environments are perceived positively or negatively by people. The
analysis is made up of a non-spatial, a spatial as well as a qualitative section. We find that neither
human perceptions nor spatial features are distributed randomly in space. Of ten features measured,
pedestrian congestion is found to be most reliably correlated with positive perceptions of the urban
environment. This study contributes to the existing work in the field by suggesting new environmental
variables to be included in future work, as well as by exploring which types of data analysis
techniques are most appropriate for architectural and urban planning research.
3
CONTENTS
1. Introduction
1.1. Motivation..................................................................................................................................................................... 8
1.2. Objective ...................................................................................................................................................................... 8
1.3. State of the art ............................................................................................................................................................. 8
1.4. Virtual Reality Experiment............................................................................................................................................ 9
1.5. Research Questions...................................................................................................................................................... 11
2. Non-spatial data analysis
2.1. Summary of method..................................................................................................................................................... 12

2.2.1. Data preprocessing..................................................................................................................................... 12
2.2.2. Data analysis................................................................................................................................................ 18
2.2. Weimar data analysis ................................................................................................................................................... 14
2.2.1. Regression analysis.......................................................................................................................................... 14
2.2.2. PCA................................................................................................................................................................... 18
2.2.3. K-means clustering analysis............................................................................................................................. 20
2.2.4. Feature extraction using decision trees.......................................................................................................... 28
2.3. Zürich data analysis ...................................................................................................................................................... 30
2.3.1. Regression analysis.......................................................................................................................................... 30
2.3.2. PCA................................................................................................................................................................... 34
2.3.3. K-means clustering analysis..........................-.................................................................................................. 37
2.3.4. Feature extraction using decision trees......................................................................................................... 42
2.3. Summary of results ...................................................................................................................................................... 47
3. Spatial data analysis
3.1. Summary of method..................................................................................................................................................... 49

3.2. Weimar data analysis ................................................................................................................................................... 50
3.2.1. Spatial autocorrelation.................................................................................................................................... 50
3.2.2. Spatially constrained clustering....................................................................................................................... 60
3.2.3. Qualitative analysis.......................................................................................................................................... 64
3.3. Zürich data analysis ...................................................................................................................................................... 71
3.2.1. Spatial autocorrelation.................................................................................................................................... 71
3.2.2. Spatially constrained clustering....................................................................................................................... 80
3.2.3. Qualitative analysis.......................................................................................................................................... 83
3.2.4. Extended qualitative analysis.......................................................................................................................... 91
3.4. Summary of results ...................................................................................................................................................... 95
4. Conclusions................................................................................................................................................................. 97
5. Bibliography............................................................................................................................................................... 99
4
LIST OF FIGURES
1. Introduction
Fig. 1.1. Devices used in the experiment............................................................................................................. 8

Fig. 1.2. Level of realness and experimentation control in real life and in VR experiments............................... 9
Fig. 1.3. Explanation of four 2D isovist measures by Benedikt, 1979, as quoted in ESUM, 2017. .................... 9
2.2. Weimar non-spatial data analysis
Fig. 2.2.1. Correlation between objective and subjective scores........................................................................... 14

Fig. 2.2.2. Scatter plot matrix of all variables, colored by subjective scores.......................................................... 15
Fig. 2.2.3. Statistically significant correlations between variables and subjective scores...................................... 16
Fig. 2.2.4. Scatter plot matrix of all variables colored by objective scores............................................................. 17
Fig. 2.2.5. Statistically significant correlations between variables and objective scores........................................ 17
Fig. 2.2.6. Elbow curve for PCA................................................................................................................................ 18
Fig. 2.2.8. Correlation plot for PCA.......................................................................................................................... 19
Fig. 2.2.10. Contributions of variables to principal components............................................................................. 19
Fig. 2.2.11. Scatter plot matrix of principal components 1-5, colored by subjective scores................................... 20
Fig. 2..2.12. Scatter plot matrix of principal components 1-5, colored by objective scores..................................... 20
Fig. 2.2.14. K = 2 clustering for Betweenness_ped vs RayPortion_obstacles.......................................................... 22
Fig. 2.2.16. K = 4 clustering for RayPortion_obstacles vs Betweenness_ped.......................................................... 24
Fig. 2.2.17. Decision tree built with all unscaled Weimar x-variables to predict subjective scores......................... 24
Fig. 2.2.18. Decision tree built with all scaled Weimar x-variables to predict objective scores.............................. 26
Fig. 2.2.19. Decision tree built with four selected variables to predict subjective scores....................................... 27
Fig. 2.2.20. Decision tree built with all variables in Weimar 360 data x-variables to predict objective scores..... 28
Fig. 2.2.21. Decision tree built with two selected variables in Weimar 360 data to predict objective scores....... 29
2.3. Zürich non-spatial data analysis
Fig. 2.3.1. Correlation between objective and subjective scores ........................................................................ 30

Fig. 2.3.2. Scatter plot matrix of all variables, colored by subjective scores.......................................................... 31
Fig. 2.3.3. Statistically significant correlations between variables and subjective scores...................................... 32
Fig. 2.3.4. Scatter plot matrix of all variables colored by objective scores .......................................................... 33
Fig. 2.3.4. Statistically significant correlations between variables and objective scores........................................ 34
Fig. 2.3.5. Elbow curve for PCA................................................................................................................................ 35
Fig. 2.3.7. Correlation plot for PCA.......................................................................................................................... 36
Fig. 2.3.9. Contributions of variables to principal components.............................................................................. 37
Fig. 2.3.10. Scatter plot matrix of principal components 1-4, colored by subjective scores.................................... 38
Fig. 2.3.11. Scatter plot matrix of principal components 1-4, colored by objective scores..................................... 39
Fig. 2.3.13. Clusters and subjective scores in k=2 clustering for principal components 3-4................................... 40
Fig. 2.3.15. Scatter plot matrix showing k=4 clustering for all scaled raw data ...................................................... 41
Fig. 2.3.16. Clusters and objective scores in k =4 clustering for all scaled raw data................................................ 42
Fig. 2.3.17. Clusters and objective scores in k=4 clustering for principal components 1-2..................................... 43
Fig. 2.3.18. Decision tree built with all x-variables to predict subjective scores....................................................... 44
Fig. 2.3.19. Decision tree built with selected variables to predict subjective scores............................................... 45
Fig. 2.3.20. Decision tree built with all-variables to predict objective scores.......................................................... 45
Fig. 2.3.21. Decision tree built with selected variables to predict objective scores................................................ 46
3.2. Weimar spatial data analysis
Fig. 3.2.1. Chroropleth maps of subjective and objective scores........................................................................... 50

Fig. 3.2.2 Chroropleth maps of principal components 1-2.................................................................................... 51
Fig. 3.2.3 Chroropleth maps of Betweenness_ped and Betweenness_car........................................................... 51
Fig. 3.2.4. Chroropleth maps of Min_radial and RayPortion_obstacles................................................................. 52
Fig. 3.2.5. Connectivity graph.................................................................................................................................. 53
5
Fig. 3.2.6 Moran’s scatter plot and pseudo p-value for objective scores.............................................................. 54
Fig. 3.2.7 Moran’s scatter plot and pseudo p-value for subjective scores............................................................ 54
Fig. 3.2.8 Moran’s scatter plot and pseudo p-value for PC1.................................................................................. 54
Fig. 3.2.9 Moran’s scatter plot and pseudo p-value for......................................................................................... 54
Fig. 3.2.10 Moran’s scatter plot and pseudo p-value for Betweenness_car............................................................ 54
Fig. 3.2.11 Moran’s scatter plot and pseudo p-value for Betweenness_ped.......................................................... 55
Fig. 3.2.12 Moran’s scatter plot and pseudo p-value for Min_radial..................................................................... 56
Fig. 3.2.13 Moran’s scatter plot and pseudo p-value for RayPortion_obstacles................................................... 56
Fig. 3.2.14 Spatial correlogram for subjective scores.............................................................................................. 57
Fig. 3.2.15 Spatial correlogram for objective scores................................................................................................ 57
Fig. 3.2.16 Spatial correlogram for principal component 1..................................................................................... 58
Fig. 2.2.17 Spatial correlogram for principal component 2.................................................................................... 58
Fig. 3.2.18 Spatial correlogram for Betweenness_car.............................................................................................. 58
Fig. 3.2.19 Spatial correlogram for Betweenness_ped............................................................................................ 59
Fig. 3.2.20 Spatial correlogram for RayPortion_obstacles....................................................................................... 59
Fig. 3.2.22 Spatially constrained k=2 clustering with using two select variables ................................................... 62
Fig. 3.2.23 Colocation maps for subjective scores and non-spatial and spatial clustering..................................... 62
Fig. 3.2.24 Spatially constrained k=4 clustering for two select raw data variables and latitude and longitude..... 63
Fig. 3.2.25 Summary map of break point locations for different variables............................................................. 64
Fig. 3.2.26 Different variables’ breakpoints in the context of a description of the video....................................... 66
Fig. 3.2.27 Subjective scores in the context of a description of the video............................................................. 67
Fig. 3.2.28 Betweenness_ped values in the context of a description of the video................................................ 68
Fig. 3.2.29 Video screenshots from locations with positive subjective scores....................................................... 69
Fig. 3.2.30 Video screenshots from locations with neutral subjective scores........................................................ 70
3.3. Zürich spatial data analysis
Fig. 3.3.1. Chroropleth maps of subjective and objective scores........................................................................... 72

Fig. 3.3.2. Chroropleth maps of principal components 1-2..................................................................................... 72
Fig. 3.3.3. Chroropleth maps of Betweenness_ped and Betweenness_car........................................................... 73
Fig. 3.3.4. Chroropleth maps of Min_radial and RayPortion_obstacles................................................................. 73
Fig. 3.3.5. Connectivity map for Zürich data when the distance threshold between neighbors is 9.0 m............. 74
Fig. 3.3.6. Moran’s scatter plot and pseudo p-value for objective and subjective scores..................................... 75
Fig. 3.3.7. Moran’s scatter plot and pseudo p-value for principal components 1-2............................................. 75
Fig. 3.3.8. Moran’s scatter plot and pseudo p-value for Betweenness_car and Betweenness_ped.................... 76
Fig. 3.3.9. Moran’s scatter plot and pseudo p-value for Min_radial and RayPortion_obstacles........................... 76
Fig. 3.3.11 Spatial correlogram for discrete objective scores. Spatial autocorrelation = 0 at c. 90 meters........... 77
Fig. 3.3.12 Spatial correlogram for subjective scores. Spatial autocorrelation = 0 at c. 76 meters........................ 77
Fig. 3.3.13 Spatial correlogram for PC1. Spatial autocorrelation = 0 at c. 30 meters............................................. 78
Fig. 3.3.14 Spatial correllogram for PC2. Spatial autocorrelation = 0 at c. 135 meters........................................... 78
Fig. 3.3.15 Spatial correlogram for Betweenness_car. Spatial autocorrelation = 0 at c. 125 meters..................... 78
Fig. 3.3.16 Spatial correlogram for Betweenness_ped. Spatial autocorrelation = 0 at c. 152 meters................... 79
Fig. 3.3.18 Spatially constrained k=2 clustering for principal components 3-4...................................................... 80
Fig. 3.3.19 Colocation maps for subjective scores and non-spatial and spatial clustering..................................... 81
Fig. 3.3.20 Spatially constrained k=4 clustering for two principal components 1-2............................................... 82
Fig. 3.3.21. Summary map of break point locations for different variables............................................................ 83
Fig. 3.3.22 Different variables’ breakpoints in the context of a description of the video....................................... 85
Fig. 3.3.23 Subjective scores in the context of the textual description of the video............................................. 86
Fig. 3.3.24. Betweenness_car in the context of the textual description of the video............................................. 87
Fig. 3.3.25 Betweenness_ped in the context of the textual description of the video........................................... 88
Fig. 3.3.26 Video screenshots from locations with positive subjective scores....................................................... 89
Fig. 3.3.27 Video screenshots from locations with neutral subjective scores........................................................ 90
Fig. 3.3.28. Simplification of subjective scores......................................................................................................... 92
Fig. 3.3.30. Simplified subjective scores and buffer typologies................................................................................ 92
Fig. 3.3.29. Simplified subjective scores and greenery ............................................................................................ 92
Fig. 3.3.31. Simplified subjective scores and facade transparency ......................................................................... 92
Fig. 3.3.32. Subjective scores and traffic speed........................................................................................................ 93
Fig. 3.3.33. Subjective scores and building construction dates................................................................................. 94
6
LIST OF TABLES
1. Introduction
Table 1.2. Explanation of the ten environmental variables in our data................................................................. 10
2.2. Weimar non-spatial data analysis

Table 2.2.7. Variance of different principal components........................................................................................... 18
Table 2.2.9. Table showing contributions (in %) of variables to relevant principal components............................. 19
Table 2.2.13. A comparison of k=2 clustering results for different data..................................................................... 22
Table 2.2.15 A comparison of k=4 clustering results for different data..................................................................... 23
2.3. Zürich non-spatial data analysis

Table 2.3.6. Amount of variance in data which different principal components account for.................................. 34
Table 2.3.8. Contributions of variables to principal components............................................................................. 35
Table 2.3.12. A comparison of k=2 clustering results for different data..................................................................... 38
Table. 2.3.14. A comparison of k=4 clustering results for different data..................................................................... 40
2.4. Summary of results of non-spatial data analysis

Table. 2.4.1. Summary of significant correlations found in the regression analysis.................................................. 47
Table. 2.4.2. Summary of which variables were found important in the clustering analysis.................................... 47
Table. 2.4.4 Differences in Betweenness_ped values for Weimar and Zürich.......................................................... 48
Table. 2.4.3 Summary of which variables were found important in the decision tree analysis............................... 48
3.2. Weimar spatial data analysis

Table 3.2.14 Summary of Moran’s I statistic for different variables........................................................................... 56
Table 3.2.21. Summary of extent of spatial autocorrelation for different variables................................................... 60
3.3. Zürich spatial data analysis

Table. 3.3.10. Summary of Moran’s I statistic for different variables........................................................................... 77
Table 3.3.17 Summary of extent of spatial autocorrelation for different variables................................................... 79
7
1. INTRODUCTION
1.1. Motivation
The livability of cities is an important topic which impacts the lives of most people in the world.
According to the United Nations (2018), 55% of the world’s population currently lives in urban areas
and by 2050, this figure is projected to increase to 68% (p. 2). In addition to land use, transportation
planning and economic structures, street-scale urban planning decisions are important for improving
urban residents’ quality of life. Firstly, attractive environments contribute to the social and psycho-
logical health of people and communities (Mehta, 2013, p. 56). In addition, according to Clifton et al
(2008), the quality of street-scale design can influence which routes people choose, which means that
understanding what constitutes high-quality design can improve transportation planning. There is
also evidence that high-quality design can make dense environments equally attractive as low-density
environments. This is important because it is widely accepted that in order to be functional and
sustainable, future cities must be dense (Clifton, Ewing, Knaap and Song, 2008, pp. 35, 38).
Another trend which is relevant to the present research is the increasing availability of data on
people’s interaction with the urban environment. Sensors, mobile data and video data combined
with modern image processing techniques provide massive amounts of information on people’s
movements in space. With appropriate analytical techniques, this data could inform urban planning
and hence make cities more livable. This study aims to further our knowledge of which methods are
most appropriate for making use of quantitative data relating to urban environments.
1.2. Objective
This study aims to increase our current understanding of how humans perceive the built environ-
ment. In particular, the study aims to determine which features of the built environment people
perceive as positive, and which features are perceived negatively. The goal of this research is to use
quantitative data to find correlations between features and perceptions.
Another goal of the study is to test whether certain commonly-used statistical methods are
appropriate for analysing data relating to urban environments. In particular, we wish to evaluate
whether the data gathered and the methods used in this study could benefit future urban planning
research or practice.
1.3. State of the Art
Existing work on human perception of the built environment is most often based on questionnaires
and field observations of people interacting with the environment. Major figures in the field, such
as Jane Jacobs, Jan Gehl and William Whyte, conclude that diverse functions, ground floor glazing,
human scale buildings, liveliness, abundance of seating and safety from traffic are key to creating
high-quality urban space (Jacobs, 1961; Gehl, 1971/2011; Whyte, 1980). More recently, Vikas Mehta
has combined existing studies with his own empirical research to identify inclusiveness, meaningful
activities, comfort, safety and pleasurability as the five dimensions of urban space quality. He divides
8
these dimensions into 42 subcomponents including, for example, opening hours of public space,
range of activities and behaviors supported by the space, climatic comfort of the space, perceived
safety from traffic, and sense of enclosure provided by the space (Mehta, 2013, p. 58-68).
There is also an emerging trend of using more technologically advanced methods to investigate the
relation between spatial features and human perceptions. Several studies have developed tools for
using big data from social media to investigate which kinds of places people find attractive (for ex-
ample Girardin et al, 2009 and Resch et al, 2015). Another approach is to use electroencephalograph
(EEG) signals and GPS trackers to measure people’s physiological responses to different locations.
Using this method in their study of 30 participants, Li et al (2016) found that isovist parameters such
as compactness, occlusivity and maximum visibility are correlated with people’s emotions, with high-
er compactness and greater visibility causing positive emotions. In other words, the study suggests
that people prefer places with relatively long vistas and clear boundaries. On the other hand, another
study with 13 participants found that high values for isovist parameters perimeter and occlusivity
predict negative perceptions of the environment (Hijazi et al, 2016). A third study investigating 30
participants’ perceptions of different locations in Wiedikon, Zürich, found that temperature, humidity
and isovist area were most strongly correlated with people’s physiological responses (ESUM, 2017).
Due to the limited number and at times contradictory findings of these studies, it is clear that more
sensor and EEG-based research on human perceptions of the built environment is still needed before
firm conclusions can be formed.
1.4. Virtual Reality Experiment
The present study follows the principles of previous studies which correlate participants’ physiological
responses to environmental features, albeit with some modifications. The first difference compared to
past studies is that participants’ responses are measured in Virtual Reality (VR) environments instead
of outdoors. The two VR environments were created by stitching 360 degree videos from Weimar,
Germany and Zürich, Switzerland, which were 19:49 and 18:20 minutes long respectively (Fig. 1.1-a).
The second difference compared to most previous studies is that in addition to physiological mea-
surements, a direct survey was used to measure participants’ responses to the urban environment.
This was done at five-second intervals in the video, and the GPS location at that moment was also
recorded. Two responses were therefore collected for each location: a “subjective” response based
on participants’ conscious evaluation of the environment, as well as an “objective” response based
on physiological measurements. Subjective ratings were recorded by asking participants to press
a clicker at each location using the scale 0—no preference, 1—positive, 2—neutral, 3—negative.
Throughout the experiment, the participant’s physiological state was also recorded using an Empatica
E4 wearable device, which measures electro-dermal activity (EDA), a measure of sympathetic nervous
system arousal (Fig. 1.1-b). Overall, data was successfully collected on 20 participants in the Weimar
experiments and 31 participants in the Zürich experiment.
a) b)
Fig. 1.1. a) 14 GoPro Hero 3+ cameras

fastenend to a remote control. These
were used to record the 3D video. b)
Empatica e4-wristband for measuring
participants’ physiological responses.
9
Using VR data has several benefits. Firstly, all participants experience the exact same environment,
making different people’s perceptions of the space more comparable with each other. Factors such as
weather and random sounds and flows of people are the same for all participants. Removal of such
confounding factors is especially important for comparing different people’s objective responses,
since they are heavily influenced by environmental conditions like temperature and light (ESUM,
2017). The second benefit of using virtual reality data is that it enables precise measurement of
environmental features (Fig. 1.4.2). Based on the video, ten features were precisely quantified. In
traditional fieldwork, these features could be described qualitatively but it would be very hard to
observe them all at once on a moment-to-moment basis – especially if the researcher must also
simultaneously observe the behavior of humans in the space
Figs 1.4.3-4 explain the environmental variables which were measured in the VR video. Most of
the variables relate to 2D or 3D isovists. An isovist is the field of view visible from a given location
(the gray areas in Fig 1.4.3.). 2D isovist variables are Area, Perimeter, Compactness, Occlusivity, and
minimum radial (Min_radial). Volume and the amount of sky and obstacles visible (RayPortion_sky
and Rayportion_obstacles) are 3D variables. Betweenness_car and Betweenness_ped measure the
amount of car and pedestrian congestion respectively.
Fig. 1.2. Abstraction of level of realness and experimenta- Fig. 1.3. Explanation of four 2D isovist measures by
tion control in real life and in VR experiments. Benedikt, 1979, as quoted in ESUM, 2017. The black points
represent people and gray areas represent the area they
can see.
Variable Explanation
Area Area of the 2D isovist (m2)
Perimeter Perimeter of the 2D isovist (m)
Compactness Ratio of the isovist’s area to perimeter in relation to a circle (a circle has the largest possible ratio, 1)
Occlusivity Length of the “open” edges in the 2D isovist
Volume Volume of the 3D isovist
Min_radial Minimum line of sight (distance between the viewing point and the closest edge of the 2D isovist)
RayPortion_obstacles Amount of obstacles visible in 3D isovist
RayPortion_sky Amount of sky visible in 3D isovist
Betweenness_car Amount of cars visible
Betweenness_ped Amount of pedestrians visible
Table 1.2. Explanation of the ten environmental variables in our data.
10
1.5. Research Questions
The research questions in this study are:

1. Which features of the urban environment are perceived as positive or negative by humans?
2. To what extent do subjective and objective scores correlate with each other?
3. Is the data gathered in this research sufficient? Does it leave out some environmental variables
which would be important for understanding human perception of the built environment?
4. Which data analysis techniques are most appropriate for this kind of study?
The first question is most important because its answer will directly expand our knowledge of urban
planning. However, answers to questions 2-4 could help future researchers improve their methodol-
ogy and perhaps make interesting findings. This is especially important because urban environmental
research using big data and physiological sensors is still very new (as discussed in Section 1.3).
The analysis is divided into two sections: non-spatial and spatial analyses. The first, non-spatial
section treats the measurements like any other data and uses techniques such as regression, principal
components analysis (PCA), clustering and decision trees to find correlations between environmental
features and human perceptions. The spatial analysis extends this discussion by studying how the
variables and perceptions relate to space/location. It also broadens the research perspective with a
qualitative analysis which takes into account environmental features for which no quantitative data
was gathered.
11
2. NON-SPATIAL ANALYSIS
2.1. Summary of method

In the non spatial analysis, statistical methods such as Principal components analysis (PCA), statistical
modeling methods like regression, and K means algorithm were applied on the collected data. The
aim is to find correlations between environmental features and objective and subjective scores. The
following operations and analyses were conducted for Weimar 360 and Zürich 360 data:
2.1.1. Data Preprocessing
Data were collected on 20 participants in Weimar and 31 participants in Zürich. Each participant gave
subjective responses at 5-second intervals throughout the experiment. On the other hand, physio-
logical response was measured continuously throughout the experiment, so that each participant’s
data constituted a single signal. Each physiological signal thus had to be quantified at a rate of five
seconds. This signal processing was conducted with Ledalab. The GPS location of the video image
was also calculated every five seconds in the video. Thus, the processed data for each participant
contained a subjective score, an objective score and a GPS location at five second intervals in the
video. This amounts to 212 data points in the case of Zürich participants and 237 data points in the
case of Weimar participants.
In the next step, the different participants’ data was merged so that each GPS location would have a
single subjective/objective score representing all participants. Different participants’ subjective scores
were combined into a discrete value using mode, while objective scores were averaged. The pro-
cessed data for Zürich and Weimar therefore consists of 212 and 237 data points respectively, each
containing a GPS location, 10 environmental variables, the mode of subjective scores and the average
objective score. A more detailed account of data preprocessing is available in ESUM (2017).
Two additional preprocessing steps were carried out for the present analysis:
1. Objective scores were transformed from continuous to discrete in the following manner:
• Original score in 0-25th percentile of all scores in the data file --> discrete score of 0
• In some cases, several data points near the cut-off point for a given percentile had the same
objective score. These data points had to be included in the same category of discrete scores.
For this reason, some discrete score categories are bigger than others.
2. Observations with subjective score = 0 were removed from analyses of correlations between
variables and subjective scores.
• Giving a subjective score of 0 is similar to answering “I don’t know” in a questionnaire. It does
not indicate a lower or higher score than subjective score 1. Including subjective scores with
value 0 would therefore interfere with analyses which rely on a linear scale in subjective values.
12
2.1.2. Data Analysis
The non-spatial data analysis consists of the following parts:
1. Scatter plots of raw data to roughly visualize correlation between variables and objective/subjec-
tive ratings:
• Scatterplot matrix to find variables whose correlation to ratings looks promising
• More detailed regression analysis for selected variables
• Regression and Parallel coordinate plot to study correlation between subjective and objective
scores
2. Principal Component Analysis (PCA) to determine whether some combination of variables is cor-
related with ratings:
• Analysis of how many principal components are relevant
• Elbow curve
• Table showing how many components explain what % of variance in the dataset
• Analysis of which variables contribute to which principal components
• Correlation Plot
• Contribution Plot
• Loadings Table
• Contributions Table
• Scatter plot matrix to visualize relationship between relevant principal components and ratings
3. K-means clustering on to check whether clusters correspond to objective / subjective ratings:

• Matrix plot showing clustering on PCs identified as important in the elbow curve (2.1)
• More detailed analysis of plots with clear clusters and how these clusters relate to subjective/
objective scores
• Measuring entropy and purity to evaluate whether clusters in PC data correspond to subjective/
objective scores
4. Feature extraction using a decision tree to investigate which variables are able to predict objec-
tive/subjective scores:
• Decision tree built using all the variables as input
• Plot illustrating the decision tree
• Significant variables identified by the decision tree
• Using the decision tree to predict objective/subjective scores for test data
• Confusion matrix to evaluate accuracy of the test
• Decision tree is built using interesting variables as input
• same process as for 4.1.
13
2.2. Weimar data analysis
2.2.1. Regression analysis
Before investigating the relationship between environmental variables and human perceptions, it is
important to understand how our two different measures of perceptions relate to one another. Fig.
2.2.1. illustrates the relationship between objective and subjective scores using a regression plot,
a parallel coordinate plot, and a table. It highlights how subjective scores fall overwhelmingly into
categories 1-2 (positive and neutral ratings), with only two observations having a value of 3 (negative
rating). According to the Figs 2.2.1-a and 2.2.1-b, there is no strong relationship between subjective
and objective scores in the Weimar dataset (Pearson’s coefficient r = 0.13 and p-value (p) = 0.048). It
is thus necessary to analyse the correlation between environmental features and human subjective
and objective perceptions independently.
a) b)
Subjective
r = 0.13 score Subjective
p = 0.048 0
1
2
3
Subjective score
Objective
continuous
Objective
discrete
Objective_discrete
c)
Fig. 2.2.1. a) Subjective vs objective score. b) Parallel coordinate plot visualising the correlation between subjective and
objective scores. Each line represents an observation. Colors represent subjective scores. The plot shows that data points
with a high/low subjective score do not generally have a correspondingly high or low objective score. c) Table showing
mean objective discrete scores for subjective scores 1-3. Data points with subjective scores 1 and 2 both have mean
objective discrete scores of 1.3, implying that there is no significant correlation between objective and subjective scores.
A first step towards understanding the relationship between environmental variables and ratings
is a scatter plot matrix, as shown in Fig 2.2.2. for subjective scores and in Fig. 2.2.4. for objective
scores. According to Fig. 2.2.2., most individual x-variables in the raw Weimar 360 data do not seem
to be correlated with subjective scores. However, the bottom row of the matrix suggests that some
relationship may exist for the variables RayPortion_obstacles and Betweenness_ped. This is because,
for these variables, the amount of data points per each value of y (subjective rating) seems to vary at
least slightly depending on the x-axis value.
14
Area
Perimeter
Compactness
Occlusivity
Min_radial
Volume
RayPortion_sky
RayPortion_obstacles
Betweenness_car
Betweenness_ped
Subjec-
tive score
Fig. 2.2.2. Scatter plot matrix showing all variables plotted against each other and subjective scores. The plots show no
clear relationship between variables and subjective scores.
Due to the inconclusiveness of the scatter plot matrix, r was determined for the correlation between
each variable and subjective scores. Figs 2.2.3. a-d) shows the variables which have statistically
significant correlations with subjective scores, along with r and the p-value. The variables which
have statistically significant correlations with subjective scores are: Betweenness_ped, RayPortion_
obstacles, Betweenness_car and Min_radial.
The same analysis was conducted for variables and objective scores (Figs 2.2.4-5). The only variable
which has a statistically significant relationships with objective scores is Betweenness_ped, so a
regression plot is only shown for this variable.
15
a) b)
c) d)
Fig 2.2.3. Selected variables vs. subjective score. The variables were chosen because of their statistically significant
correlation with subjective scores. Figures and values were obtained using Pearson’s method.
a) Betweenness_ped vs Subjective score. r = -0.38, p = 3.2*10 - 9

b) RayPortion_obstacles vs Subjective score. r = 0.18, p = 0.0055
c) Betweenness_car vs Subjective score. r = 0.17, p = 0.0082
d) Min_radial vs Subjective score. r = -0.16, p = 0.017
16
Area
Perimeter
Compactness
Occlusivity
Min_radial
Volume
RayPortion_sky
Betweenness_car
Betweenness_ped
Objective
score
Fig 2.2.4. All x-variables plotted against each other. The bottom row shows the objective score plotted against all x-vari-
ables. Objective scores are represented by colors. The plots show no clear relationship between objective scores and
x-variables.
Fig 2.2.5. Betweenness_ped vs. objective score. It is the only variable with a statistically significant correlation with
objective scores, with r = -0.22 and p = 0.00079
17
2.2.2. PCA
Since the correlations between individual variables and objective/subjective scores are not very
strong, a PCA is conducted to investigate whether some combination of x-variables shows a stronger
relationship.
First, an elbow curve is drawn to determine the number of relevant principal components (PC)
(Fig. 2.2.6). The contributions of different components is also shown in Table 2.2.7. These results
suggest that only principal components 1-5 should be considered in the analysis of Weimar_360
data, since altogether these account for 94,9% of variance in the x-variables of the data. Within the
principal components, PC1 is clearly the most important with a contribution of 52%, compared to a
contribution of 15% for PC2.
Fig 2.2.6. Elbow curve showing how much of variance in data different principal components account for..
Table 2.2.7. Variance in data which different principal components account for..
The next step is to determine what the different principal components mean. Fig. 2.2.8 and Table
2.2.9 contain information on how much each x-variable contributes to the different principal
components. Based on this analysis, out of the 10 variables, only Compactness, Betweenness_car and
Betweenness_ped do not have a big influence on the variance in the data. Although these variables
contribute somewhat significantly to PC2-3, those principal components account only for 15% and
13% of variance in the data respectively, as shown in Table 2.2.7, making the impact of these variables
on overall variance quite small. Fig. 2.2.10 presents a more detailed view of the contributions of and
relationship between different variables in PC1-2.
18
Fig 2.2.8. Correlation plot showing which variables are important in different principal components in Weimar data.
Table 2.2.9. Table showing contributions (in %) of variables to relevant principal components.
Fig 2.2.10. Contributions of variables to different principal components. In this case the contribution is calculated in a
different way compared to Fig/Table 2.2.8-9.
19
Finally, how do the principal components relate to subjective and objective scores? Figs 2.2.11-12
show the relevant principal components plotted against one another with data points colored
according to objective and subjective scores. There does not appear to be a clear relationship
between any component and the ratings, although it is possible that some correlation exists between
PC4 or PC5 and subjective scores.
Fig. 2.2.11. Principal components 1-5 plotted

against each other. Subjective scores are
represented by colors. The plots suggest that PC5,
and to a lesser extent PC4, may correlate with
subjective scores.
Fig. 2.2.12. Principal components 1-5 plotted

against each other. Objective scores are
represented by colors. The plots do not suggest
any correlations between principal components
and objective scores.
20
2.2.3. K-means clustering analysis
The scatter plot matrix on PC data (Fig. 2.2.11) suggests that there may be a relationship between
PC4 and PC5 and subjective scores. Clustering is performed to investigate this and other possible
relationships further. Table 2.2.13. shows a comparison of clustering results for different selections
from Weimar data. In all cases, all 233 observations were used (subjective score zeroes were
excluded), but the x-variables used for clustering differ. The idea is to search for a clustering method
which differentiates subjective scores 1 and 2 neatly into two different clusters. Two clusters are
suited to this data since subjective scores fall overwhelmingly into categories 1-2. The success of the
clustering can be judged by the mean subjective score in each cluster, and by the external cluster
validation metrics entropy and purity (Kent State University, Sauwen). Ideally, one cluster would have
a mean score of 1 and the other 2, entropy would be 0 and purity would be 1. Within cluster sum of
squares and between/total sum of squares measure cluster cohesion and independence, however
they considered less important metrics for cluster quality in this case.
Table 2.2.13 shows that, judging by the metric of difference in mean subjective scores, the most
successful clustering is based on the variables Betweenness_ped and RayPortion_obstacles. In this
case, cluster 1 has a mean subjective score of 1.27, while cluster 2 has a mean subjective score of
1.56. This clustering also narrowly beats all the others on the metrics of entropy and purity (both
0.62). The clustering is presented in greater detail in Fig. 2.2.14. The clusters are differentiated both
along the x- and y-axes, indicating that the both variables contribute to the clusters’ difference in
mean subjective scores. This suggests that both variables are correlated with subjective scores.
This same procedure is applied to objective scores, albeit with four centers used for clustering to suit
the four categories of objective_discrete scores. Table 2.2.15 shows the performance of clustering
on different datasets. This time, standard deviation is used as a measure of the difference between
objective score values in different clusters. According to the analysis, clustering based on PC4-5 and
on the variables Betweenness_ped and RayPortion_obstacles performs best (standard deviation
in cluster mean objective scores = 0.39). Clustering based on the two original data variables is
chosen for further illustration in Fig. 2.2.16. This shows that clusters 1-2, which have the greatest
difference in mean objective score, are differentiated along both x-and y-axes. This suggests that both
RayPortion_obstacles and Betweenness_ped correlate with objective scores. Entropy for clustering
based on two variables is 0.97, while purity is 0.32.
Clustering with k=2 performed significantly better than that with k=4. This is not surprising, since
with fewer clusters and fewer distinct score values, there is less chance of getting scores and clusters
mixed up. In both cases, clustering with RayPortion_obstacles and Betweenness_ped performed
best. It is interesting to compare this result with the regression study in section 2.2.1, which found
that RayPortion_obstacles, Betweenness_ped, Betweenness_car and Min_radial have statistically
significant relationships with subjective scores. Clustering based on these four variables (column 4
in Tables 2.2.13 and 15) performed worse than the clustering based on only Betweenness_car and
Min_radial. For objective scores, it is less surprising that these two variables resulted in the best
clustering since they had the most significant relationships with objective scores in the regression
analysis.
21
Scaled raw
Scaled raw data,
Scaled raw data, 2 select 4 select PC1, PC1, PC2, PC4,
Raw data data variables variables PC1-5 PC2 PC5 PC3 PC5
Cluster 1:
frequency of
subjective score
values
1 36 58 79 37 58 79 78 92 44
2 37 38 29 37 38 58 57 61 30
3 1 0 0 1 0 2 2 1 2
Cluster 2:
frequency of
subjective score
values
1 100 78 57 99 78 57 58 44 92
2 58 57 66 58 57 37 38 34 65
3 1 2 2 1 2 0 0 1 0
Mean
Subjective score
Cluster 1 1,53 1,40 1,27 1,52 1,40 1,45 1,45 1,41 1,45
Cluster 2 1,38 1,45 1,56 1,38 1,45 1,39 1,40 1,46 1,41
Mean of all data

1,42 1,42 1,42 1,42 1,42 1,42 1,42 1,42 1,42
Difference
between Cluster
1-2 means 0,15 -0,05 -0,29 0,14 -0,05 0,05 0,05 -0,05 0,03
Within cluster
sum of squares
Cluster1 1,11*e^13 721,46 157,0416 263,05 661,98 401,61 248,74 123,3557 118,9886
Cluster2 1,51*e^13 724,16 153,5338 348,86 666,11 303,06 206,56 184,9175 87,31174
Between_ss/
total_ss 85,5 % 37,7 % 33,1 % 34,1 % 39,7 % 55,1 % 65,6 % 36,5 % 49,8 %
External cluster
validation
Entropy 0,65 0,65 0,62 0,65 0,65 0,65 0,65 0,66 0,65
Purity 0,59 0,58 0,62 0,58 0,58 0,58 0,58 0,58 0,58
Table 2.2.13. A comparison of k=2 clustering results for different data. The table investigates how well clusters align with
subjective scores. This is measured by the frequency and mean of scores in different clusters, as well as by entropy and
purity. In addition, cluster quality is measured by within cluster sum of squares and the ratio of between to total sum
of squares. There is almost no difference in how well the clusters align with subjective scores (measured by entropy
and purity). While the distribution of subjective scores in each cluster changes depending on the data used, the mean
subjective scores for both clusters are also quite similar for all data sets. The best clustering is outlined in red (chosen
because it has the greatest difference in mean subjective scores between clusters)
Fig 2.2.14 Betweenness_ped vs

RayPortion_obstacles. K-means
clustering for these variables with
k=2. Colors represent clusters, while
shapes represent subjective scores. The
frequencies and means of subjective
scores in each cluster are shown on
the top right.. In the figure, the clusters
are differentiated along both the x- and
y-axes, implying that both variables are
correlated with subjective scores.
22
Scaled raw Scaled
data, 4 raw data,
Scaled raw select 2 select
Raw data data variables variables PC1-5 PC1, PC2 PC1, PC5 PC2, PC3 PC4, PC5
Cluster 1: frequency
of objective score
values
0 16 8 29 1 19 10 18 27 32
1 8 9 26 3 17 2 11 31 20
2 6 7 23 3 19 3 13 26 15
3 8 11 21 6 18 3 15 32 11
of objective score
values
0 9 17 7 14 17 18 15 18 23
1 8 7 8 11 7 17 21 7 22
2 7 6 2 7 6 19 17 7 17
3 12 8 0 3 8 18 23 6 17
of objective score
values
0 25 19 19 26 26 33 21 12 3
1 25 17 8 26 25 31 16 11 3
2 23 19 11 23 18 23 12 10 9
3 22 18 20 26 23 32 14 16 12
of objective score
values
0 20 26 15 29 8 9 16 13 12
1 17 25 16 18 9 8 10 9 13
2 14 18 14 17 7 5 8 7 9
3 18 23 19 25 11 7 8 6 20
Mean
objective score
Cluster 1 1,16 1,60 1,36 2,08 1,49 0,94 1,44 1,54 1,06
Cluster 2 1,61 1,13 0,71 0,97 1,13 1,51 1,63 1,03 1,35
Cluster 3 1,44 1,49 1,55 1,49 1,41 1,45 1,30 1,61 2,11
Cluster 4 1,43 1,41 1,58 1,43 1,60 1,34 1,19 1,17 1,69
Mean of all data

1,42 1,42 1,42 1,42 1,42 1,42 1,42 1,42 1,42
Standard deviation in
means 0,16 0,17 0,35 0,39 0,17 0,22 0,16 0,25 0,39
Within cluster
sum of squares
Cluster 1 1,09*e^12 1,09*e^12 137,88 9,48 289,10 34,03 33,15 89,14 28,36
Cluster 2 1,14*e^12 1,14*e^12 45,19 41,15 255,47 100,28 66,03 51,13 25,67
Cluster 3 3,15*e^12 3,14*e^12 110,22 43,26 339,00 213,22 54,16 26,66 15,07
Cluster 4 2,83*e^12 2,82*e^12 129,25 49,10 123,24 42,16 38,73 50,53 17,65
Between_ss/
total_ss 95,5 % 95,5 % 55,4 % 69,8 % 56,6 % 75,7 % 85,7 % 67,9 % 73,7 %
External cluster
validation
Entropy 0,99 0,99 0,96 0,97 0,98 0,98 0,98 0,98 0,96
Purity 0,31 0,31 0,32 0,32 0,31 0,30 0,33 0,33 0,37
Table 2.2.15. A comparison of k=4 clustering results for different data. The table investigates how well clusters align
with objective scores. This is measured by the frequency and mean of scores in different clusters, as well as by entropy
and purity. In addition, cluster quality is measured by within cluster sum of squares and the ratio of between to total
sum of squares. The dataset outlined in red appears to give the best clustering result.
23
Fig. 2.2.16. RayPortion_obstacles vs Betweenness_ped. k=4 clustering for data with only these two variables. More
information on the clustering is available in column 4 in Table. 2.2.15 (outlined in red). The data points’ shapes are
determined by objective scores, while colors represent clusters.
2.2.4. Feature extraction using decision tree
A decision tree is a simple supervised learning method in which the computer tries to learn the
relationship between a given y-variable and one or more x-variables. Here, the rpart function is used
to create a decision tree for predicting subjective and objective scores based on 1) all x-variables in
the data and 2) x-variables which have statistically significant relationships with subjective/objective
scores. The decision-making model is first trained on a portion of the data (training data) and pruned.
To see if it works well, the model is then applied to a new dataset (test data) to predict subjective or
subjective scores. In addition, the model is applied to Zürich data to see if the same variables predict
objective/subjective scores in Zürich as in Weimar. The training dataset is based on the Weimar 360
data. Separate trees are trained on scaled and unscaled data. The Weimar training and test sets for
subjective predictions contain 179 and 59 data points respectively, while the Zürich test dataset
contains 210 data points. For objective scores, the Weimar training and test sets contain 178 and 160
data points respectively, while the Zürich test set contains 213 data points.
Fig. 2.2.17 shows a decision tree which uses all 10 x-variables in unscaled Weimar data to predict
subjective scores. The variables identified as important in the tree are shown in part b) in the figure.
The decision tree weighs Betweenness_ped and Betweenness_car most heavily when trying to
predict objective scores. A large Betweenness_ped value is on its own enough to predict a subjective
score of 1 (root split); after that the tree also considers Betweenness_car, then Perimeter, Area and
finally Betweenness_ped again. The prediction accuracy is 72,8% when applied to Weimar test data
and 69.5% when applied to Zürich data. The confusion matrices showing predicted vs. actual results
are shown in the figure’s parts d-e for Weimar and Zürich data respectively.
24
a)
b)
c)
d) e)
Fig. 2.2.17. a) Decision tree built with all unscaled Weimar x-variables to predict subjective scores. b) Importance of vari-
ables in the decision tree. c) Errors of the decision tree at each split. d) Confusion matrix for the tree built when applied to
Weimar 360 test dataset. For test data, the tree has an accuracy of 72,8%. e) Confusion matrix for the tree when applied
to the entire unscaled Zürich 360 dataset. For predictions on Zürich data, the tree has an accuracy of 69,5%.
25
Fig 2.2.18 shows the same tree as in Fig 2.2.17 based on all raw data variables to predict subjective
scores, however this time the training and test data have been scaled. Strangely, the scaling massively
decreases the accuracy of the tree’s predictions for Zürich data. When the training data and Zürich
data are scaled, the tree’s accuracy is 28.2% when predicting Zürich scores, compared to 69.5% with
unscaled data.
Another decision tree for predicting subjective scores is shown in Fig 2.2.19. Instead of taking all
x-variables of the Weimar_360 data into account when predicting objective scores, this tree uses
only the variables Betweenness_ped, RayPortion_obstacles, Betweenness-car and Min_radial.
These variables were chosen because they showed statistically significant correlation with subjective
scores in Section 2.2.1. As in the previous tree, the two most important variables in this tree are
also Betweenness_car and Betweenness_ped. Unlike before, in this tree, Betweenness_car is given
more importance than Betweenness_ped. A large Betweenness_ped value is on its own enough
to predict a subjective score of 1 (root split), but after that the tree bases most of its decisions on
Betweenness_car. RayPortion_obstacles is used in only one split, while Min_radial is not used at all.
With an accuracy of 88.1%, this tree performs better than the previous one at predicting subjective
scores for Weimar data (Figs. 2.2.18-19 d). On the other hand, when applied to Zürich data, its
prediction accuracy of 60% is worse than for the previous tree (Figs. 2.2.18-19 e).
a)
b) c)
Fig. 2.2.18. a) Decision tree built with all scaled Weimar x-variables to predict objective scores. b) The confusion matrix
and accuracy of the tree when applied to Weimar data is also the same as in Fig 2.2.17. c) However, when the tree is
applied to scaled Zürich data, its performance worsens dramatically. The accuracy of the scaled tree is 23.8%, compared
to the unscaled tree’s 69.5%.
26
a)
b)
c)
d) e)
Fig. 2.2.19. a) Decision tree built with four selected variables in Weimar 360 data to predict subjective scores. The
variables are Betweenness_car, Betweenness_ped, RayPortion_obstacles and Min_radial (chosen because they were
statistically significantly correlated with subjective scores). b) Importance of variables in the decision tree. c) Errors of the
decision tree at each split. d) Confusion matrix for the tree when applied to the Weimar 360 test dataset. For test data,
the tree has an accuracy of 88,1%. e) Confusion matrix for the tree when applied to the entire Zürich 360 dataset. For
predictions on Zürich data, the tree has an accuracy of 60,0%.
27
The decision trees are less successful at predicting objective scores. First, a decision tree taking into
account all x-variables to predict objective scores is shown in Fig 2.2.20. The tree was forcibly grown
so that it would make predictions for all objective score values. Using the standard pruning method,
in which branches are pruned unless they decrease the error in Fig 2.2.20-c, resulted in a tree with
zero splits. Indeed, the table shows that the x-error of the tree only increases as branches are added,
which means that the tree would perform better if it predicted the same value all the time. When
applied to Weimar data, the tree’s accuracy is 28%, while accuracy for predictions in Zürich is 15%. In
other words, the tree performs equally badly as – or even worse than – chance.
a)
b)
c)
d) e)
Fig. 2.2.20. a) Decision tree built with all x-variables in Weimar 360 data to predict objective scores. b) Importance of
variables in the decision tree. c) Errors of the decision tree at each split. d) Confusion matrix for the tree when applied
to the Weimar 360 test dataset. For test data, the tree has an accuracy of 28,3%. e) Confusion matrix for the tree when
applied to the Zürich 360 dataset. For predictions on Zürich data, the tree has an accuracy of 15,5%.
28
Another decision tree for predicting objective scores, built with variables Betweenness_ped and
RayPortion_obstacles, is shown in Fig 2.2.21. Again, standard pruning procedure would have reduced
the tree to a single root. The tree was grown so that it would make at least one prediction for
each category. Its accuracy in predicting Weimar and Zürich objective scores is 21.7% and 28.2%
respectively. Again, this is not much different from chance.
The decision tree analysis supports the conclusion of previous analyses that Betweenness_ped is the
variable most highly correlated with subjective scores. Contrary to the regression analysis in section
2.2.1., the decision tree considers Betweenness_car a better predictor of subjective scores than
RayPortion_obstacles. The decision tree does not prioritise Min_radial. Another conclusion is that
objective scores are difficult to correlate with any variables. The decision tree analysis diminishes the
importance of the correlations between Betweenness_ped and RayPortion_obstacles and objective
scores in section 2.2.1., since using these variables to predict scores works as badly as chance.
a)
b)
c)
d) e)
Fig. 2.2.21. a) Pruned decision tree built with two selected variables in Weimar 360 data to predict objective scores. The
variables are Betweenness_car and RayPortion_obstacles (chosen because they were statistically significantly correlated
with objective scores). b) Importance of variables in the decision tree. c) Errors of the decision tree at each split. d)
Confusion matrix for the tree when applied to the Weimar 360 test dataset. For test data, the tree has an accuracy of
21,7%. e) Confusion matrix for the tree when applied to the Zürich 360 dataset. For predictions on Zürich data, the tree
has an accuracy of 28,2%.
29
2.3. Zürich data analysis
The procedure described in section 2.1. is also followed for Zürich 360 data. In the following
discussion of results, the method will be explained in less detail than before.
2.3.1. Regression analysis
Fig 2.3.1. illustrates the relationship between subjective and objective scores for Zürich data.
Subjective and objective scores show even less correlation than in Weimar, with r=0.02 and p = 0.78
in the regression analysis. The mean objective_discrete score for data points with subjective score=1
is 1.42, while that of data points with subjective score=2 is 1.46. Again, this means that correlations
between variables and human perceptions must be sought separately for objective and subjective
scores.
a) b)
r = 0.13
p = 0.048
Subjective score
Objective_discrete
c)
Fig. 2.3.1. a) Subjective score vs objective discrete score. The correlation is statistically insignificant and weak (r = 0.02
and p = 0.78). b) Parallel coordinate plot to illustrate how subjective scores (top) are linked to continuous objective
scores (middle) and discrete objective scores (bottom). Each line represents one data point and each color represents a
subjective score. c) Median and mean and distribution of objective scores for each category of subjective scores. Mean
and median objective scores are very similar for subjective scores 1-2.
Fig 2.3.2. shows a scatter plot matrix between original x-variables with points colored according to
subjective scores. A regression analysis was conducted to test the correlation of each variable with
subjective scores. Fig 2.3.3. shows Betweenness_ped and Min_radial – the only variables statistically
significantly correlated with subjectives scores – plotted against subjective scores. Betweenness_ped
has a correlation coefficient of 0.22 with a p-value of 0.0013, while for Min_radial r = -0.16 and p =
0.022.
Figs. 2.3.4-5 follow the same procedure for objective scores. For Zürich data, a greater number of
variables are statistically significantly correlated with objective scores than subjective ones. The
variables Betweenness_ped, Min_radal, Occlusivity, Volume, RayPortion_sky, Area and Betweenness_
car all have statistically significant correlations, although correlation coefficients are small (the
maximum value is r = -0.33 for Betweenness_car).
Again, it is noteworthy that Betweenness_ped and Min_radial showed up as statistically significant for
both subjective and objective scores. However, the direction of correlation for Betweenness_ped was
different, with r=0.22 for subjective scores and -0.3 for objective ones.
30
Area
Perimeter
Compactness
Occlusivity
Min_radial
Volume
RayPortion_sky
Betweenness_car
Betweenness_ped
Subjective
score
Fig 2.3.2. All x-variables plotted against each other. The bottom row shows the subjective score plotted against all
x-variables. Subjective scores are represented by colors. The plots show no clear relationship between subjective scores
and x-variables.
Fig 2.3.3. a-b. Selected variables vs. subjective score. The variables were chosen because they have a statistically signifi-
cant relationship with subjective scores. Figures and values were obtained using Pearson’s method.
a) Betweenness_ped vs Subjective score. r = 0.22, p = 0.0013
b) Min_radial vs Subjective score. r = -0.16, p = 0.022
31
Area
Perimeter
Compactness
Occlusivity
Min_radial
Volume
RayPortion_sky
Betweenness_car
Between-
ness_ped
Subjective
score
Fig 2.3.4. All x-variables plotted against each other. The bottom row shows the objective score plotted against all
x-variables. Objective scores are represented by colors. The plots show no clear relationship between objective scores and
x-variables.
32
a) b)
c) d)
e) f)
g)
Fig 2.3.4. Selected variables vs. subjective score. The variables were chosen because they have a statistically significant
relationship with subjective scores. Figures and values were obtained using Pearson’s method.
a) Betweenness_ped vs Objective score. r = -0.3, p = 6.2*10- 6
b) Min_radial vs Objective score. r = -0.14, p = 0.047
c) Occlusivity vs Objective Score. r = -0.14, p = 0.049
d) Volume vs Objective score. r = -0.15, p = 0.026
e) RayPortion_sky vs Objective score. r = -0.15, p = 0.026
f) Area vs Objective score. r = - 0.15, p = 0.034
g) Betweenness_car vs Objective score. r = - 0.33, p = 1.2 * 10- 6
33
2.3.2. PCA
Fig. 2.3.5 shows an elbow curve to determine the number of relevant principal components for the
Zürich_360 data. The contributions of different components is also shown in Table 3.3.6. These results
suggest that only principal components 1-4 should be considered in the analysis of Zürich data, since
altogether these account for 89,9% of variance in the x-variables of the data. PC1 is clearly the most
important principal component with a contribution of 56,5%, compared to PC2’s contribution of
17,7% PC2.
Fig. 2.3.7 and Table 2.3.8 show the contribution of each x-variable to the different principal
components. Based on this analysis, out of the 10 variables, only Min_radial and Occlusivity do not
have a big influence on the variance in the data. The most important variables in the data are Volume,
Perimeter and RayPortion_sky, all contributing more than 12% to PC1 (although Area, Compactness,
Occlusivity and RayPortion_obstacles are not far behind either, each having a contribution of >10%).
Fig. 2.3.9. shows the relation between variables in PC1-2 in more detail.
Figs. 2.3.10-11 show the relevant principal components plotted against one another with data points
colored according to objective and subjective scores respectively. There does not appear to be any
relationship between any component and the ratings, though it is possible that some relationship
exists between PC3 and subjective scores.
Fig 2.3.5. Elbow curve showing how much of variance in data different principal components account for.
Table 2.3.6. Table showing how much of variance in data different principal components account for.
34
Fig 2.3.7. Correlation plot showing which
variables are important in different principal
components in Zürich data.
Table 2.3.8. Table showing contributions (in %)

of variables to relevant principal components.
Fig 2.3.9. Contributions of variables to different

principal components. In this case the contribu-
tion is calculated in a different way compared
to Fig/Table 2-3.7-8.
35
Fig. 2.3.10. Principal components 1-4
plotted against each other. Subjective
scores are represented by colors.
The plots suggest that PC3 may be
correlated with subjective scores.
Fig. 2.3.11. Principal components 1-4

plotted against each other. Objective
scores are represented by colors. The
plots do not suggest any correlations
between principal components and
objective scores.
36
2.3.3. K-means clustering
Table 2.3.12. shows k=2 clustering performed on different datasets with the aim of finding a method
which separated subjective scores into two clear categories. The clustering is most successful when
performed on PC3-4: in this case, cluster 1 has a mean subjective score of 1.44 while cluster 2 has a
mean subjective score of 1.73 (difference = 0.29). Entropy in this case is 0.88 and purity is 0.69. Fig.
2.3.13 illustrates the result of this clustering. The clusters are more differentiated along the PC3 axis,
implying that PC3 variables are more correlated with subjective scores than PC4 variables. According
to Table 2.3.8, the most important variables in PC3 are Min_radial, Occlusivity and RayPortion_
obstacles with contributions of 22.4%, 17.4% and 12.8% respectively. It is somewhat surprising that
this clustering, which does not emphasise Betweenness_ped, performed better than the clustering
based on Betweenness_ped and Min_radial (column 3 in Table 3.2.12).
Table 2.3.14 presents a similar study for objective scores. In this case, the best clustering is based
on PC1-PC2 data, yielding a standard deviation of 0.5 for mean objective scores in different clusters.
Cluster mean scores are 1.25, 0.61, 1.40 and 2.03, while entropy and purity are 0.98 and 0.37
respectively. Clustering on unscaled raw data yields an even greater standard deviation, however this
is largely because its fourth cluster contains only 3 data points. Since objective scores categories are
all roughly the same size, this clustering is unlikely to reveal real correlations between variables and
objective scores.
Figs 2.3.15-16 present k=4 clustering performed on all raw data, which also resulted in a standard
deviation of 0.5 in mean objective scores of different clusters. Again, two clusters (1 and 3) have an
especially large difference in mean objective scores between them (means 0.62 and 2.0 respectively).
From the matrix plot in Fig 2.3.16, it is possible to find the plot with the greatest distance between
categories 1 and 3. This appears to be RayPortion_obstacles vs Betweenness_car, which is shown in
greater detail in Fig.2.3.16. Clusters 1 and 3 are extremely differentiated along the Betweenness_car
axis, suggesting that Betweenness_car is correlated with objective scores. There is much less
differentiation along the RayPortion obstacles axis.
Figure 2.3.17 presents the k=4 clustering performed on PC1-2. Clusters 2 and 4 have the largest
difference in mean objective scores (having mean scores 0.61 and 2.0 respectively) and interestingly,
these two variables are also highly differentiated along the PC2 axis of the plot. This implies that
the variables contributing most to PC2 correlate with objective scores. According to Table 2.3.8, the
greatest contributors to PC2 are Betweenness_car and Betweenness_ped, with contributions of
32.1% and 30.5% respectively.
To summarise, clustering for Zürich data suggests that Min_radial, Occlusivity and RayPortion_
obstacles are most correlated with subjective scores. Both k=4 clustering analyses performed suggest
that Betweenness_car is correlated with objective scores. The PC1-2 clustering gives an almost equal
importance to Betweenness_ped.
37
Scaled Scaled
raw data, raw data,
Scaled 2 select 6 select PC1, PC2, PC2, PC3,
Raw data raw data variables variables PC1-4 PC2 PC4 PC3 PC4
Cluster 1:
frequency of
subjective
score values
1 2 25 12 25 47 24 54 54 30
2 25 48 13 45 90 48 79 78 24
3 0 0 0 0 0 0 0 0 0
Cluster 2:
frequency of
subjective
score values
1 70 47 60 47 25 48 18 18 42
2 113 90 125 93 48 90 59 60 114
3 0 0 0 0 0 0 0 0 0
Mean
Subjective
score
Cluster 1 1,93 1,66 1,52 1,66 1,66 1,67 1,59 1,59 1,44
Cluster 2 1,62 1,66 1,68 1,66 1,66 1,65 1,77 1,77 1,73
Mean of all
data 1,66 1,66 1,66 1,66 1,66 1,66 1,66 1,66 1,66
Difference
between
Cluster 1-2
means 0,31 0,00 -0,16 0,00 0,00 0,01 -0,17 -0,18 -0,29
Within
cluster
sum of
squares
Cluster1 1,99*e^13 561,64 31,53 340,74 470,36 370,86 128,19 203,69 89,09
Cluster2 8,83*e^14 623,82 227,00 379,58 500,45 269,28 109,62 133,35 113,05
Between_ss/
total_ss 70,4 % 43,4 % 38,2 % 42,6 % 48,2 % 58,5 % 51,3 % 42,6 % 39,0 %
External
cluster
validation
Entropy 0,89 0,93 0,92 0,93 0,93 0,93 0,90 0,90 0,88
Purity 0,66 0,66 0,66 0,66 0,66 0,66 0,66 0,66 0,69
Table 2.3.12. A comparison of k=2 clustering results for different data. The table investigates how well clusters
align with subjective scores. This is measured by the frequency and mean of scores in each cluster, as well as by
entropy and purity. In addition, cluster quality is measured by within cluster sum of squares and the ratio of between
to total sum of squares. The PC3-4 dataset outlined in red appears to give the best clustering result (although unscaled
raw data has a larger difference between cluster means, this dataset was not chosen because the clusters are of very
uneven size). For column 3, the two variables are Betweenness_ped and Min_radial.
38
Fig 2.3.13. K-means clustering for PC3-4 data with k=2. Data points with subjective score = 0 have been removed from the
data used for this clustering. Colors represent clusters while shapes represent subjective scores.
39
Scaled Scaled
raw data, raw data,
Scaled 6 select 2 select PC1, PC2, PC2, PC3,
Raw data raw data variables variables PC1-4 PC2 PC4 PC3 PC4
Cluster 1:
frequency of
objective score
values
0 26 7 7 23 20 20 14 11 5
1 26 11 5 14 21 20 8 15 3
2 24 18 6 21 19 19 2 11 3
3 16 27 4 11 10 9 1 12 0
Cluster 2:
frequency of
objective score
values
0 14 19 30 6 7 13 16 14 16
1 7 13 26 6 11 7 16 8 15
2 2 13 18 13 18 2 14 2 15
3 1 14 14 22 27 1 10 1 10
Cluster 3:
frequency of
objective score
values
0 18 12 7 7 19 18 9 12 15
1 17 6 11 7 13 12 7 8 10
2 26 2 18 6 13 13 12 12 11
3 33 1 26 5 14 14 10 5 11
Cluster 4:
frequency of
objective score
values
0 1 20 14 22 12 7 19 21 22
1 0 21 9 24 6 12 20 20 23
2 0 19 10 12 2 18 24 27 23
3 2 10 8 14 1 28 31 34 31
Mean
objective score
Cluster 1 1,33 0,62 1,32 1,29 1,271429 1,25 0,60 1,49 0,82
Cluster 2 0,58 1,27 1,18 2,09 2,031746 0,61 1,32 0,60 1,34
Cluster 3 1,79 2,03 2,02 1,36 1,372881 1,40 1,61 1,27 1,38
Cluster 4 2,33 1,37 1,29 1,25 0,619048 2,03 1,71 1,73 1,64
Mean of all
data
1,46 1,46 1,46 1,46 1,46 1,46 1,46 1,46 1,46
Standard
deviation in
means 0,64 0,50 0,33 0,34 0,50 0,50 0,43 0,42 0,30
Within cluster
sum of squares
Cluster 1 2,07*e^13 136,961 63,38 19,27 208,60 121,67 16,66 22,16 7,57
Cluster 2 2,75*e^11 341,465 187,94 7,08 72,47 84,01 17,53 23,02 23,10
Cluster 3 1,42*e^13 78,7865 91,93 31,75 269,98 40,60 20,67 35,96 26,59
Cluster 4 1,09*e^10 258,371 132,81 34,97 60,29 57,56 56,04 68,34 40,48
Between_ss/
total_ss 98,8 % 61,5 % 62,6 % 78,0 % 67,9 % 80,7 % 77,4 % 74,7 % 70,6 %
External cluster
validation
Entropy 0,9447 0,94 0,96 0,96 0,94 0,93 0,96 0,95 0,98
Purity 0,3521 0,37 0,36 0,36 0,37 0,37 0,34 0,35 0,31
Table. 2.3.14. A comparison of k=4 clustering results for different data. The table investigates how
well clusters align with objective scores. This is measured by the frequency and mean of scores in
each cluster, as well as by entropy and purity. In addition, cluster quality is measured by within cluster
sum of squares and the ratio of between to total sum of squares. The PC1-2 dataset outlined in red
appears to give the best clustering result (although unscaled raw data has a larger standard deviation
between cluster means, this dataset was not chosen because the clusters are of very uneven size).
40
Fig 2.3.15. K-means clustering for scaled raw data with k=4. Colors represent clusters.
Fig. 2.3.16. RayPortion_sky vs Betweenness_car - a closer look at one of the plots in Fig 2.3.15. Colors represent clusters
while shapes represent objective scores. Mean objective scores for each cluster are shown. In this data set, clusters 1 and
3 have the largest difference in mean scores and in this plot, the clusters are highly differentiated along the x-axis. This
implies that Betweenness_car is correlated with objective scores.
41
Fig. 2.3.17. PC1 vs PC2, showing k-means clustering on PC1 and PC2 data with k=4. Colors represent clusters while shapes
represent objective scores. Mean objective scores for each cluster are shown. In this data set, clusters 2 and 4 have the
largest difference in mean scores and in this plot, the clusters are highly differentiated along the x-axis and to a lesser
extent differentiated along the y-axis. This implies that PC2 could be correlated with objective scores.
2.3.4. Feature extraction using decision trees
Again, a decision tree is created which predicts subjective and objective scores based on 1) all
x-variables in the data and 2) the x-variables identified as most important by regression analyses. The
training and test datasets are based on the Zürich 360 data, but the trees are additionally tested on
the entire Weimar data. For subjective scores, the Zürich training datasets contain 157 and 53 data
points respectively, while the Weimar test set contains 233 points. For objective scores, the Zürich
training and test datasets contain 159 and 54 data points respectively, while the Weimar test set
contains 238 points.
Fig. 2.3.18. shows a decision tree which uses all 10 x-variables in unscaled Zürich data to predict
subjective scores. As in the case of Weimar, the first node in the tree tests for Betweenness_ped. If
Betweenness_ped is below 2030, the tree predicts a subjective score of 2. Otherwise, the tree goes
on to evaluate Betweenness car, with values larger than 5*106 indicating a subjective score of 1. The
final level in the tree tests for Area. When applied to Zürich test data, the prediction accuracy of the
tree is 84.9%, but for Weimar data it is only 42.9% - worse than chance. The problem is that the tree
almost always (218 out of 233 times) ends up predicting subjective score =2 for the Weimar data. This
is probably because the Weimar data generally fails the Betweenness_ped test in the first branch of
the tree, leading the tree to predict a score of 2 without even considering other variables.
Fig. 2.3.19 shows another tree built to predict subjective scores, this time using the variables
Betweenness_ped, Betweenness_car, RayPortion_obstacles and Min_radial. These variables were
chosen because they show statistically significant correlations with subjective scores for Weimar data.
The idea is to test whether using these variables would improve the accuracy of the tree’s predictions
for Weimar data. Surprisingly, this tree performs even worse than the one based on all variables, with
prediction accuracy 79.2% for test Zürich data and 41.6% for Weimar data.
42
a)
b)
c)
d) e)
Fig. 2.3.18. a) Decision tree built with all Zürich 360 x-variables to predict subjective scores. b) Importance of variables
in the decision tree. c) Errors of the decision tree at each split. d) Confusion matrix for the tree when applied to Zürich
360 test data. For test data, the tree has an accuracy of 84,9%. e) Confusion matrix when the tree is applied to the entire
Weimar 360 dataset. For predictions on Weimar data, the tree has an accuracy of 42,9%.
43
a)
b)
c)
d) e)
Fig. 2.3.19. a) Decision tree built with selected Zürich 360 x-variables to predict subjective scores. The variables used
were Betweenness_ped, Betweenness_car, RayPortion_obstacles and Min_radial. b) Importance of variables in the
decision tree. c) Errors of the decision tree at each split. d) Confusion matrix for the tree when applied to Zürich 360 test
data. For test data, the tree has an accuracy of 79,3%. e) Confusion matrix when the tree is applied to the entire Weimar
dataset. Here, the tree has an accuracy of 41,6%.
Fig. 2.3.20. shows a decision tree which uses all 10 x-variables in Zürich data to predict objective
scores. As usual, the first node in the tree tests for Betweenness_ped. If Betweenness_ped is greater
than 94000, the tree predicts an objective score of 3. Otherwise, the tree goes on to evaluate Area
and Perimeter. When applied to Zürich test data, the prediction accuracy of the tree is 31.5%, but for
Weimar data it is only 22.2% - again, worse than chance.
Fig. 2.3.21. shows a decision tree which uses six x-variables in Zürich data to predict objective
scores. The variables chosen are Betweenness_ped, RayPortion_sky, Area, Occlusivity, Volume and
Min_radial. The variables were chosen because they show statistically significant correlation with
objective scores in the Zürich dataset. Again, the first node in the tree tests for Betweenness_ped,
which also makes up the third level of the tree, with Area being the tree’s only other variable. The
tree results in an accuracy of 27.8% for Zürich data and 18.5% for Weimar data.
44
a)
b)
c)
d) e)
Fig. 2.3.20. a) Decision tree built with all Zürich 360 x-variables to predict objective scores. b) Importance of variables
in the decision tree. c) Errors of the decision tree at each split. d) Confusion matrix for the tree when applied to Zürich
360 test data. For test data, the tree has an accuracy of 31,5%. e) Confusion matrix when the tree is applied to the entire
Weimar 360 dataset. For predictions on Weimar data, the tree has an accuracy of 22,2%.
45
a)
b)
c)
d) e)
Fig. 2.3.21. a) Pruned decision tree built with selected Zürich 360 x-variables to predict objective scores. The variables
used were Betweenness_ped, RayPortion_sky, Volume, Occlusivity, Area and Min_radial. b) Importance of variables in the
decision tree. c) Errors of the decision tree at each split. d) Confusion matrix for the tree when applied to test data from
the Zürich 360 dataset. For test data, the tree has an accuracy of 27,8%. e) Confusion matrix when the tree is applied to
the entire Weimar 360 dataset. For predictions on Weimar data, the tree has an accuracy of 18,5%.
This decision tree analysis once again supports the conclusion that Betweenness_ped is correlated
with subjective scores and probably objective scores as well, and that objective scores are very hard
to predict. Building a decision tree trained on Zürich data using variables correlated with subjective
scores in Weimar data also resulted in the conclusion that the main difference between the datasets
is not which variables are important, but rather which values those variables have. The decision
tree built on variables suited to Weimar data did very poorly at predicting subjective scores in the
Weimar experiment, probably because the values of Betweenness_ped are very different in Weimar
compared to Zürich, on which the tree was trained.
46
2.4. Summary of results
How do these different analyses help us to answer our research question on the relationship between
x-variables and subjective/objective scores? First of all, according to the regression analysis in sections
2.2.1 and 2.3.1, objective and subjective scores are not related to each other in either Weimar or
Zürich datasets. This casts doubt on the possibility of finding variables which are correlated with
both scores. The different analyses offer somewhat contradictory insights into which variables are
most linked to objective and subjective scores. However, it seems that Betweenness_ped certainly
has some correlation with scores. Min_radial has a slight negative correlation with all except Weimar
objective scores. Betweenness_ped is more strongly correlated with both datasets’ objective and
subjective scores. This is shown in Table 2.4.1. The only discrepancy is that the correlation is positive
for Zürich subjective scores and negative for all other scores.
Table 2.4.2. summarises the results of the k-means clustering analysis, showing how often each
variable was identified as important in the most successful clustering for a particular dataset and
score type. For example, if using PC1-2 data leads to the clusters most able to differentiate between
subjective scores, the important variables would be identified by examining the spread of the
clusters along the principal components and then identifying the most important contributors to
the component along which clusters were most differentiated. Nevertheless, it is important to note
that clustering did not perform very well, especially for objective scores. For the best k=4 clusters
in both Weimar and Zürich, only 2 clusters had significantly different mean objective scores and
entropy levels were close to 1. For subjective scores, 0.3 was the largest difference in clusters’ mean
subjective scores in each dataset. Entropy and purity were each around 0.6 for Weimar and 0.9 and
0.7 respectively for Zürich.
Weimar, Weimar, Zürich, Zürich, Table. 2.4.1. Summary

subj. score obj. score subj. score obj. score of results of regression
Subjective score 0,13 analysis showing
Area -0,15 statistically significant
Perimeter correlations between
Compactness variables and objective/
Occlusivity -0,14 subjective scores in each
Min_radial -0,16 -0,16 -0,14 dataset. The figures
Volume -0,15 represent r-values.
RayPortion_sky -0,15 Negative correlations
RayPortion_obstacles 0,18 are highlighted in red,
Betweenness_car 0,17 -0,33 while positive ones are
Betweenness_ped -0,38 -0,22 0,22 -0,30 highlighted in green.
Table. 2.4.2. Summary of

Weimar, Weimar, Zürich, Zürich,
results of the clustering
subj. score obj. score subj. score obj. score
analysis. The table
Area
shows which variables
Perimeter
are important in the
Compactness
most successful cases
Occlusivity x
of k-means clustering.
Min_radial x
Betweenness_ped and
Volume
RayPortion obstacles are
RayPortion_sky
emphasised because
RayPortion_obstacles x x
they were identified as
Betweenness_car x
important more than
Betweenness_ped x x x
once..
47
Table 2.4.3. summarises the results of the decision tree analyses. The analysis shows that predicting
objective scores is impossible: no prediction for these scores performed much better than chance.
It is thus wiser to focus on subjective scores. The analysis suggests that Betweenness_car and
Betweenness_ped are the most important for predicting subjective scores. This was determined
by selecting the most accurate decision tree for each prediction and listing the two most important
variables for each tree.
Given that Betweenness_car and Betweenness_ped are quite good at predicting subjective scores for
both Zürich and Weimar data, one might think that trees trained on one of these datasets could also
predict subjective scores in the other. However, this is not the case, especially for a tree trained on
Zürich data. (Table 2.4.3). To understand why, it might help to consider the Betweenness_ped splits
which occur in subjective root of subjective Zürich and Weimar decision trees. In the Weimar tree, a
data point is predicted to have subjective score 1 if Betweenness_ped is greater than 5828. In Zürich,
a Betweenness_ped score of less than 2030 leads to a predicted score of 2. Both splits suggest that
a high Betweenness_ped is linked to positive scores (subjective score 1); however, there is a large
difference between these scores. Table 2.4.4. shows that in Weimar, Betweenness_ped values are
greater than in Zürich by all metrics considered. Scaling the decision tree’s training and test data is
attempted in Fig. 2.2.21; however, this only worsened the Weimar tree’s ability to predict subjective
scores in Zürich. The reasons for the two datasets’ different relationship between Betweenness_ped
and subjective score are discussed in the spatial analysis of Zürich.
Despite some contradictions, comparing the results of the analyses allows us to make some tentative
conclusions on which variables are important. Regression, clustering and decision trees all suggest
that Betweenness_ped is an important variable with relation to scores. These analyses also suggest
that Betweenness_car and, to a lesser extent, RayPortion_obstacles are important . In the PCA,
these values are more important in Zürich data than in Weimar data. Even if the variables are not
important in the PCA, however, they can still have the most significant correlations with objective and
subjective scores. The analysis also suggests that subjective scores are much more clearly linked to
environmental features than objective ones, which perform badly in regression, clustering and decision
tree analyses.
Weimar, Weimar, Zürich, Zürich, Table. 2.4.3. Summary of the re-
subj. score obj. score subj. score obj. score sults of the decision tree analyses.
Area
The top (yellow) part of the table
Perimeter 11
Compactness 20
shows the variable importance of
Occlusivity the two most important variables
importance
Variable
Min_radial in each tree. The green part states

Volume the cutoff value and variable in
RayPortion_sky
the first split of the tree and what
Betweenness_car 34 11 29 26 score the tree predicts based on
Betweenness_ped 41 30 26 this. The blue part of the figure
Betw_ped Compactness Betw_ped Betw_car states the prediction accuracy of
First split >= 5828 >= 0.47 < 2030 >=94 e^3 each tree for predicting test data
yes --> 1 yes --> 0 no --> 2 no --> 3 and the opposite data set’s data.
Accuracy in
predicting
Prediction accuracy
own data set's

test data 88,14 % 28,33 % 84,91 % 31,48 %
Accuracy in
predicting
other data set's
test data 60,00 % 15,49 % 42,92 % 22,22 %
Figure Fig. 2.2.19 Fig. 2.2.20 Fig. 2.3.18 Fig. 2.3.20
Betweenness_ped
Zürich Weimar Difference
Table. 2.4.4. Differences in Betweenness_ped values for Weimar
Average 1979 4 831 -2 852 and Zürich. All metrics for Betweenness_ped indicate higher values
Min 650 1 140 -490 for Weimar data. This could have implications for the predictive
Max 4028 10 778 -6 750 power of decision trees when applied to data sets which they have
Median 1868 4 498 -2 630 not been trained on.
48
3.SPATIAL ANALYSIS
3.1. Summary of method
This study adds a spatial dimension to our previous analysis by investigating the spatial distribution of
features and human perceptions. The analysis is conducted with GeoDa, using methods described in
the GeoDa Workbook (Anselin, 2018). The spatial analysis contains the following parts:
1. Spatial autocorrelation: quantifying the influence of location on attributes

• Exploratory choropleth maps to visualize whether the following variables are spatially correlat-
ed, and where break points in the data are located
• discrete objective score
• subjective score
• PC1
• PC2
• Betweenness_ped
• Betweenness_car
• RayPortion_obstacles
• Moran’s scatter plot & Moran’s I
• Spatial correlogram to investigate similarity of values across longer distances
2. Spatially constrained k-means clustering

• Clustering is performed in the same way as in Chapter 2, but taking into account location
• Colocation maps are drawn to evaluate correlation of k=2 clusters with subjective scores
3. A qualitative analysis examining the results of (1) and (2) in the context of the ESUM videos
• The video is summarised in text form and combined with various maps
• Investigation of what happens in the video at important breaks in the data, e.g. when subjective
scores turn neutral after being positive for a long time
• Search for new variables which might be correlated with subjective scores
4. For Zürich data only, an extended qualitative analysis examining subjective score maps in the con-
text of a previous study of the environmental features of the area
• The previous study provides precise maps of features such as vegetation, traffic speeds, building
construction dates, traffic buffer zones and facade types
• Investigation of whether subjective scores seem to be correlated with the features mapped in
the previous study
49
3.2. Weimar data analysis
3.2.1. Spatial autocorrelation
Spatial autocorrelation is a statistical measure of the similarity of attributes in nearby locations. In this
section, Moran’s I is computed to investigate how similar different data points are to their immediate
neighbors. A spatial correlogram is used to extend the analysis of spatial autocorrelation to longer
distances.
The first step is to understand the influence of location on variables through a set of choropleth
maps where data points are colored according to their values for a given variable. Eight variables
are investigated: objective and subjective scores, PC1, PC2, Betweenness_car, Betweenness_ped,
Min_radial and RayPortion_obstacles. Figs 3.2.1 a-b. show the distribution of objective and subjective
scores on the map. Fig 3.2.2 a-b. show the distribution of PC1 and PC2 on the map. The principal
components are computed from the x-variables of the original data in the same way as in the
previous analysis. In this map, their values are divided into four quantiles. Figs 3.2.3 a-b show the
distribution of Betweenness_car and Betweenness_ped on the map.. Finally, Figs 3.2.4 a-b show the
distribution of Min_radial and RayPortion_obstacles on the map.
a) b)
Fig. 3.2.1. Choropleth map of a) subjective and b) objective scores for Weimar 360 data.
50
a) b)
Fig. 3.2.2 Choropleth map of a) PC1 and b) PC2 for Weimar 360 data.
a) b)
Fig. 3.2.3 Choropleth map of a) Betweenness_ped and b) Betweenness_car values for Weimar 360 data.
51
a) b)
Fig. 3.2.4. Choropleth map of a) Min_radial and b) RayPortion_obstacles values for Weimar 360 data.
Figs 3.2.1-4 suggest that neighboring locations tend to have similar values, both in the case of ratings
and physical features. The western part of the Weimar area is especially homogenous. Another
interesting observation is that often breaks in the values (a sudden jump from a high to a low value
or vice versa) occur at street intersections. Previous studies have also found that the strongest
emotional responses to environments often occur as one enters a new space, often after crossing an
intersection (see for example Hijazi, 2016, p. 9). Out of the variables analysed, objective scores seem
to be least correlated to neighboring locations’ scores although there, too, data points in a quartile
are likelier to be beside points in a neighboring quartile rather than a more distant one.
Setting spatial weights

Before spatial autocorrelation can be calculated, spatial weights must be set for the data points. This
is because evaluating whether similar locations have similar attributes requires data on which data
points are considered to be near each other. Spatial weights provide this data: a data point near the
current point will have a high spatial weight, while a very faraway data point will have a spatial weight
of zero. A spatial weight matrix is created which contains the weights between all the points in the
data.
In our case, it makes sense to set distance-band weights. This means that points i and j are considered
neighbors whenever j falls within a critical distance of i. The critical distance is usually the smallest
distance at which all points have at least one neighbor. If the density of points varies greatly in the
data, some points will have very many neighbors while others have only one. This can be problematic
when calculating spatially lagged variables or local spatial autocorrelation. In our case, however, the
data points are very evenly distributed, and it is easy to find a critical distance which creates two
neighbors for all data points (except for the points at the start and end of the city tour, which each
have only one neighbor). Fig 3.2.5. shows the connectivity graph for Weimar data obtained from our
spatial weights.
52
Fig. 3.2.5. Connectivity graph for Weimar 360 data.
Each point is connected to another point if the distance
between them is less than 8.0 m (the critical distance in
our case). As a result, all points except the end points in
the South have exactly two neighbors.
Moran’s I calculation
Once spatial weights are set, spatial autocorrelation can be calculated by comparing a point’s
attribute value to the average value of that attribute among the data point’s neighbors. The average
value of the neighbors is known as the spatially lagged value. For example, if data point A has an
objective score of two and its neighbors have objective scores of three and four, the spatially lagged
objective score for point A is 3.5. Figure 3.2.6-a shows a Moran scatter plot of spatially lagged
objective scores vs. original objective scores. Moran’s I is a measure of spatial autocorrelation, with
higher values indicating greater similarity between data points in similar locations. In our case the
slope is 0.27, which means that points near each other are modestly similar.
Next, a “pseudo p-value” is calculated to ensure that our positive Moran’s I finding is not a result
of mere chance. This is done by creating a randomized reference distribution for the statistic. In
other words, the values for objective scores and spatially lagged scores are randomly distributed
on the map 999 times and spatial autocorrelation is calculated for each permutation. If the random
permutations often result in an equal or larger Moran’s I value than the actual value, that means the
original Moran’s I is probably not statistically significant. The threshold for significance is 0.05 like with
the ordinary p-value, as long as a very large number of permutations is used (usually 999 is enough).
Fig 3.2.6-b shows the result of the pseudo p-value calculation for objective scores. The purple lines
represent frequencies for different Moran’s I values. The thick green line is the actual Moran’s I for
our data. Despite 999 permutations, the reference distribution never resulted in an equal or larger
value than the observed Moran’s I, indicating that it is very unlikely that our Moran’s I is obtained due
to chance.
Figs 3.2.7-2.2.13 show Moran’s scatter plots and pseudo p-value diagrams for subjective scores, PC1,
PC2, Betweenness_car, Betweenness_ped and Min_radial and RayPortion obstacles respectively. The
results are summarised in Table 3.2.14. In short, all attributes except for Objective_discrete show very
strong spatial autocorrelation, ranging from 0.75 for subjective scores to 0.97 for Betweenness_ped.
The results are also all significant. The randomized reference distributions always yields smaller
spatial autocorrelations than the ones that actually exist. Since this statistic compares attributes of
locations which are =< 8m apart from one another, it is not very surprising that high levels of spatial
autocorrelation are found.
53
a) b)
Fig 3.2.6 Spatial autocorrelation for objective scores.

a) Moran’s scatter plot and the resulting Moran’s I. Moran’s I is 0.27.
b) Pseudo p-value calculation for the Moran’s I. The pseudo p-value is 0.001.
a) b)
Fig 3.2.7 Spatial autocorrelation for subjective scores.

a) b)
Fig 3.2.8 Spatial autocorrelation for PC1.

54
Fig 3.2.9 Spatial autocorrelation for PC2.
Fig 3.2.10 Spatial autocorrelation for Betweenness_car.

Fig 3.2.11 Spatial autocorrelation for Betweenness_ped.

55
Fig 3.2.12 Spatial autocorrelation for Min_radial.
Fig 3.2.13 Spatial autocorrelation for RayPortion_obstacles.

Table 3.2.14 Summary of spatial autocorrelation analysis results for all data variables and PC1-2. These correlations are
recorded for data points which are at most 8,0 m apart from one another.
56
It is interesting to know how far data points can lie from one another while still maintaining spatial
correlation. The spatial correlogram can help to accomplish this. Its idea is to compute correlations
for all pairs of data points on the map, to measure the distance between the points in each pair, and
to regress the correlations against the distances to compute spatial autocorrelation. No weight matrix
is used.
Figs. 3.2.14-16 show correlograms for the main variables under analysis: objective and subjective
scores, PC1, PC2, Betweenness_car, Betweenness_ped and Volume respectively. The histogram’s bins
represent the frequency of data point pairs with a given distance between them. The statistics below
the histogram provide more detailed information on the lower and upper (min-max) bounds of each
bin, as well as the number of pairs for each bin. The line above the histogram represents the amount
of spatial autocorrelation for pairs with a given distance between them. Fig 3.2.14., for instance,
shows that that spatial autocorrelation reaches zero at the dark purple bin. Again, the statistics in the
bottom provide the exact distance at which correlation is zero: ~0.327 km.
Table 3.2.21 summarises the results of the correlogram analysis. Spatial autocorrelation extends
furthest for the subjective score (327 m) and least for Compactness (57). Correlation for objective
and subjective scores as well as Betweenness_car and Betweenness_ped extends significantly
further than for the other variables. On the other hand, the slope for the objective score is shallow
and approaches zero very gradually from its modest starting point of I=0.27. Perhaps in reality, the
distance over which spatial autocorrelation has an effect is smaller.
Fig 3.2.14 Spatial correlogram for subjective scores.

Spatial autocorrelation = 0 at c. 327 meters.
Fig 3.2.15 Spatial correlogram for objective scores.

Spatial autocorrelation = 0 at c. 176 m.
57
Fig 3.2.16 Spatial correlogram for PC1. Spatial autocor-
relation = 0 at c. 62 m.

relation = 0 at c. 92 m.
Fig 3.2.18 Spatial correlogram for Betweenness_car.

58
Fig 3.2.19 Spatial correlogram for Betweenness_ped.
Fig 3.2.20 Spatial correlogram for

RayPortion_obstacles. Spatial autocorrelation = 0 at c.
57 m.
59
Table 3.2.21. Summary of how far positive spatial autocorrelation extends for different variables and PC1-2.
3.2.2. Spatially constrained clustering
Spatially constrained clustering differs from regular clustering by taking into account location. Instead
of clustering together observations whose values are similar, spatially constrained clustering groups
together observations whose values and locations are similar. There are several ways to perform
spatially constrained clustering: 1) including spatial variables in k-means clustering, 2) clustering
with a skater algorithm based on a minimum spanning tree which reflects the contiguity structure of
the observations and 3) using a max-p regions model (Anselin, 2017). The most suitable clustering
type for our purposes seems to be k-means with spatial variables included. This is because our data
consists of points instead of regions, and clustering methods 2-3 involve the use of contiguity-based
weights which cannot be created for points.
Fig 3.2.22 shows spatially constrained k=2 clustering based on variables Betweenness_ped and
RayPortion_obstacles plus location variables. The first two variables are chosen because clustering
based on them performs best in our comparison of clusterings in section 2.2.3. The location variables
added to this clustering are Latitude and Longitude. The spatial clustering is able to differentiate
subjective scores better than the non-spatial clustering. Average subjective scores are 1.21 and 1.85
for clusters 1 and 2 respectively, creating a difference of 0.64 points between clusters. This compares
to a difference of 0.29 points in non-spatial clustering.
A colocation map provides a convenient way to evaluate the success of clustering. Fig 3.2.23
compares spatially constrained clustering results with the results of clustering based on just
Betweenness_ped and RayPortion_sky and no spatial variables. Including spatial variables significantly
improves clustering performance: 77.7% of data points are classified accurately compared to just
61.3% for ordinary clustering. 53 points are misclassified, but 7 of these have subjective scores of 0 or
3 and had no chance of being classified correctly. The clustering algorithm has particular difficulty in
classifying points in regions where subjective score values change frequently.
60
a)
b)
Fig 3.2.22. a) K-means clustering with k=2 for two select raw data variables and latitude and longitude. The variables
are Betweenness_ped and RayPortion obstacles. The ratio of between to total sum of squares is 0.28. b) Distribution of
subjective scores in each cluster.
61
Accuracy = 185/238 = 77,7%
Accuracy = 146/238 = 61,3%
Fig 3.2.23 Colocation map for clustering results and subjective scores a) when spatial variables are used in clustering and
b) when only Betweenness_ped and RayPortion_sky are used. The maps show how often the data points both belong
to cluster 2 and have a subjective score of 2, or belong to cluster 1 and have a subjective score of 1. Gray data points do
not have the same value for the cluster and subjective score. Five and two data points with subjective scores 0 and 3
respectively are also included in this data and account for some of the gray points
62
Fig 3.2.24 presents a similar analysis for objective scores. Again, the non-spatial variables used are
RayPortion_obstacles and Betweenness_ped because clustering on them perfumed best in section
2.2.3. This time, adding spatial variables to the clustering does not increase standard deviation
between cluster means compared to non-spatial clustering. Comparing the clustering map with the
map of objective scores also does not show that clusters correspond closely to particular scores. This
was a likely outcome given that objective scores are so weakly correlated with both non-spatial or
spatial variables.
a)
b)
Fig. 3.2.24 a) K-means
clustering with k=4 for two
select raw data variables and
latitude and longitude. The
variables are Betweenness_ped
and RayPortion_obstacles. The
ratio of between to total sum of
squares is 0.59. b) Distribution
of objective score values in the
clusters.
63
3.2.3. Qualitative analysis
Studying environmental features exclusively through 2D maps and ten numeric variables obviously
misses out on many other qualities of the environment, and may lead to erroneous conclusions. It
is therefore useful to include a qualitative, broader description of the environmental qualities into
our analysis. This is done by a textual summary of the qualitative contents of the video in different
locations on the map. The most interesting results of the spatial study (such as which areas on the
map display high subjective scores) can then be analysed in the context of a broader, qualitative
description of environmental features.
Fig. 3.2.25 summarises interesting locations for different variables. The figure was obtained by
analysing break points in the choropleth maps in section 3.2.1. For instance, if there has been a long
period of low Betweenness_ped and suddenly the value becomes higher for another significant
interval, the point where change happens would be identified as a break point. Places where values
changed very frequently were not considered break points. Fig 3.2.25 shows that a few areas on
the map contain break points for many different variables. These are often intersections. This makes
sense, because often different streets have different attributes.
Fig 3.2.26 analyses these breakpoints in the context of a textual description of the video. The textual
description was made with no knowledge of where different break points were located. The number
of points covered by one text box was determined by the similarity of the environment in those
points – if the description in the text box no longer held, then a new text box was created. It is thus
interesting to see that in many cases, the break points between text boxes concur with break points in
the different variables. This suggests that the variables do capture changes in the environment which
are meaningful to a human observer. Beyond this and the importance of intersections, no trend was
found for what causes break points.
Fig. 3.2.25 A summary map of interesting locations for differ-

ent variables. The squares highlight places where a variable’s
values undergo an important shift, such as suddenly going
from a very low to a very high value. Places where values
change very frequently were not marked, however. The
background map shows subjective scores.
64
Fig 3.2.27 analyses the location of subjective scores in the context of the video. Positive subjective
scores (score =1) are focused on the southern part of the region, and in this region the text boxes also
contain many more mentions of trees, pedestrian areas and frequency of people. These factors seem
most correlated with positive subjective scores, while for example shopping streets are more often
found also in areas perceived as neutral. It seems that shops on their own, unless accompanied by
people, do not increase the attractiveness of urban environments.
Fig. 3.2.28 presents a similar analysis contextualisation for Betweenness_ped values. The idea is to
test whether the qualitative perception of liveliness correlates with Betweenness_ped values. Out
of 6 mentions of “many people”, 1 was located in an area of upper 4th quartile Betweenness_ped
values, 3 in upper 3rd quartile areas, 1 in a 2nd quartile area, and one in an area split between second
and fourth quartiles. All three mentions of “no people” in the northern part of the region are located
with bottom quartile Betweenness_ped values. Other text boxes in the northern region say “few” or
“not many” people, further reinforcing the link between human perceptions of Betweenness_ped
and the quantitative values in the data. Perhaps there is slightly less agreement for high values of
Betweenness_ped because for example the sounds that people make in the video create a feeling of
liveliness, even if no actual people are seen due to them being in e.g. a café. The computer might not
pick up on this and may therefore underestimate the number of people.
Figs 2.3.39-30 present a sample of screenshots from places with consistently positive and neutral
subjective scores respectively. The main difference between these sets of images seems to be
the difference in amount of people, with places with neutral scores having almost no pedestrians.
Although far from an exhaustive study, this suggests that trees, shops and historical buildings on their
own do not create attractive urban environments if they are not populated with other people.
65
Street with light traffic,
Wider few pedestrians. 1 shop
sidewalk, then front, otherwise
plaza with buildings seem
trees forming a residential. Facades are
Street with car road, sidewalk and shop lively. Historical Descending stairs, then
canopy. Road
fronts. Cars and some busses passing monumental building sidewalk next to
passing next to
frequently. Small trees in between side across road. Tall, white elevated bushes/grass.
plaza has very
walk and car road. Not many people. modern building ahead No people. 2-3 storey
light traffic. No
Historic building across street.(05:50- at the end of the street. historical buildings,
people. (07:40-
07:35, 70-91). (08:15-09:40, 99-116). white modern building.
08:10, 92-98).
Facades not particularly
Street surrounded by buildings with open. Light traffic.
shop facades on ground floor. Car (09:45-10:35, 117-127).
road and sidewalk. Two buses pass by.
(05:20-05:40, 64-68). Larger windows in buildings. Narrow
sidewalk and road. A few shops, no
Wide open space with buildings trees. No people. Light traffic. (10:40-
and trees in the background. 11:55, 128-143).
Pedestrian area with car road
nearby (no cars though). Few Wider street, historical
people. (05:00-05:15, 60-63) buildings with shops on
ground floor. A few trees and
pedestrians. Light traffic.
Same plaza with trees. Now, more
(12:00-13:15, 144-159).
traffic beside plaza. One bus drops off
lots of people right in front of the
Street with lots of shops,
camera. (04:20-04:55, 51-59)
parked cars and some
people. Light traffic. Then
Plaza with lots of trees forming a
wider pedestrian area with
canopy. Car road beside plaza but
terraces and trees. A few
no traffic, 1 parked car. (03:30-
trucks parked. (13:20-14:45.
04:15, 38-51)
160-177).
Pedestrian shopping street with lots

Wide pedestrian plaza with café
of people. Ground floor facades have
terraces, a church, historical
lots of glazing. Buildings not as historic
architecture. Many pedestrians.
as before. (02:15-03:25, 27-41)
(14:50-15:15, 178-183)
Large pedestrian plaza with outdoor

terrace. Some pedestrians, fewer Shopping street with a narrow
trees than before. Historic buildings. car road. A few pedestrians and
(01:25-02:10, 16-26) parked cars. Historical and
modern buildings mixed. (15:20-
Pedestrian street with trees, 16:30, 184-198)
historical buildings, an outdoor
terrace. Many people. Historic
Shopping street with many
buildings. (00:00:-01:20, 0-15)
people, opens up to a plaza
with a market and historical
Street plaza opening up to a wide pedestrian architecture such as arches. A
boulevard with a tree canopy. Lots of people. few trees. (16:35-17:45, 199-
Old buildings. Benches where people can sit. 213)
(18:25-19:49, 221-237)
Pedestrian shopping street with historical
buildings, quite many pedestrians. Contains
plants and trees. (17:50-18:20, 214-220) 68
Fig. 3.2.26 Different variables’ breakpoints in the context of a textual description of the video. Variables’ breakpoints
often concur with breakpoints in the textual description (points where one text box changes to another). Colors in the
description correspond to colors of the breakpoints The exceptions to this are the green and red texts, which represent
trees and shops respectively. These variables did not exist in the raw data, so they are not included in the break point 66
map, either.
Street with car road and sidewalk. Shop lively. Historical Descending stairs, then
canopy. Road
fronts, cars and some busses passing monumental building sidewalk next to
passing next to
plaza has very
light traffic. No
people. (07:40-
08:10, 92-98).
Same plaza with trees. Now, more pedestrians. Light traffic.
traffic beside plaza. One bus drops off (12:00-13:15, 144-159).
camera. (04:20-04:55, 51-59) Street with lots of shops,
04:15, 38-51)
160-177).

as before. (02:15-03:25, 27-41)
(14:50-15:15, 178-183)

buildings. (00:00:-01:20, 0-15)
(18:25-19:49, 221-237)
plants and trees. (17:50-18:20, 214-220) 69
Fig. 3.2.27 Subjective scores in the context of a textual description of the video. Places with positive subjective scores
(=1) often occur in places where the textual description mentions many trees, many people and pedestrian areas. Some
mentions of small trees or few people are not highlighted in the textual description.
67
Street with car road, sidewalk and shop lively. Historical Descending stairs, then
canopy. Road
fronts. Cars and some busses passing monumental building sidewalk next to
passing next to
plaza has very
light traffic. No
people. (07:40-
08:10, 92-98).
Same plaza with trees. Now, more pedestrians. Light traffic.
traffic beside plaza. One bus drops off (12:00-13:15, 144-159).
camera. (04:20-04:55, 51-59) Street with lots of shops,
04:15, 38-51)
160-177).

as before. (02:15-03:25, 27-41)
(14:50-15:15, 178-183)

buildings. (00:00:-01:20, 0-15)
(18:25-19:49, 221-237)
70
plants and trees. (17:50-18:20, 214-220)
Fig. 3.2.28 Betweenness_ped values in the context of a textual description of the video. Places with high Betweenness_
ped values (lots of pedestrian congestion) are generally also described as having many people in the textual description.
68
Fig. 3.2.29 Screenshots from frames at 00:00, 01:00, 02:00 and 03:00 in the
Weimar video (from top). During the first 3 minutes and 35 seconds of the
video, locations received consistently positive subjective ratings (subjective
score = 1)..
69
Fig. 3.2.30 Screenshots at 6:00, 7:00, 8:00 and 9:00 minutes in the Weimar
video (from top). From 6:00-9:30, subjective scores were consistently neutral
(with one negative rating in between).
70
3.3. Zürich data analysis
3.3.1.Spatial autocorrelation
Figs 3.3.1-4 show choropleth maps for objective and subjective scores, PC1, PC2, Betweenness_car,
Betweenness_ped, Min_radial and RayPortion_obstacles respectively. The figures again suggest
that neighboring locations tend to have similar values, both in the case of ratings and physical
features. The North-Eastern and South-Eastern regions in this dataset receive consistently positive
ratings, while ratings for other areas are almost exclusively neutral. Interestingly, the highest values
for Betweenness_ped are located in the neutral regions. Subjective score, Betweenness_car,
Betweenness_ped and PC2 tend to be similar along the entire length of a street, changing abruptly at
intersections. PC2’s most important contributors are Betweenness_car and Betweenness_ped, so the
similarity of these variables’ behavior is not surprising.
Fig 3.3.5. shows the connectivity graph for Zürich data obtained from our spatial weights. In this case,
the distance threshold for neighbors is 9.0 m. All observations have at least two neighbors, while one
data point at the intersection of Aemtlerstrasse and Kalkbreitstrasse has 6 neighbors. This threshold
value was chosen because even a slightly smaller one would have resulted in isolated points with
no neighbours. The uneven number of neighbors suggests that the data points are more unequally
distributed than in the Weimar data.
Spatial autocorrelation is calculated for the Zürich variables in the same manner as for Weimar.
Figures 3.3.6-9 show Moran scatterplots and accompanying pseudo p-value graphs for PC1, PC2,
Betweenness_car, Betweenness_ped and Min_radial and RayPortion obstacles respectively.
Results are summarised in Table 3.3.10. Again, objective score is the variable with least spatial
autocorrelation (0.32). The range of values for spatial autocorrelation is overall slightly lower than in
Weimar, ranging from 0.65 for Occlusivity to 0.94 for Betweenness_car, compared to a range of 0.75
to 0.97 for Weimar data. Nevertheless, all results are statistically significant. Again, these findings are
not very surprising given that the statistic compares attributes of locations which are <= 9 m apart
from one another.
Figs 3.3.11-16 show correlogram plots for objective and subjective scores, PC1, PC2, Betweenness_
car and Betweenness_ped respectively. Results of the correlogram analysis are presented in Table
3.3.17. Spatial autocorrelation extends furthest for the Betweenness_ped (152 m) and least for
Perimeter and Compactness (25 and 27 m respectively). Strangely, according to this analysis,
spatial autocorrelation for objective scores extends further than for subjective scores (90 and 76 m
respectively). A more meaningful metric for the present analysis is clearly the magnitude of spatial
autocorrelation, not how far it extends.
71
a) b)
Fig. 3.3.1. Map of Zürich data with points colored according to their a) objective and b) subjective scores.
a) b)
Fig. 3.3.2. Map of Zürich data with points colored according to their a) PC1 and b) PC2 values.
72
a) b)
Fig. 3.3.3. Map of Zürich data with points colored according to their a) Betweenness_car and b) Betweeness_ped values.
a) b)
Fig. 3.3.4. Map of Zürich data with points colored according to a) Min_radial and b) RayPortion_obstacles values.
73
Fig. 3.3.5. Connectivity map for Zürich data when the distance threshold between neighbors is 9.0 m.
74
a)
b)
Fig. 3.3.6. Moran scatter plot for a) objective and b) subjective scores. The pseudo p-value graph is shown on the right.
a)
b)
Fig. 3.3.7. Moran scatter plot for a) PC1 and b) PC2.
75
a)
b)
Fig. 3.3.8. Moran scatter plot for a) Betweenness_car and b) Betweenness_ped.
a)
b)
Fig. 3.3.9. Moran scatter plot for a) Min_radial and b) RayPortion_obstacles.
76
Table. 3.3.10. Summary of Moran’s I statistic for Zürich 360 data.
Fig 3.3.11 Spatial correlogram for discrete objective

scores. Spatial autocorrelation = 0 at c. 90 meters.
Fig 3.3.12 Spatial correlogram for subjective scores.

Spatial autocorrelation = 0 at c. 76 meters
77
relation = 0 at c. 30 meters.
Fig 3.3.14 Spatial correllogram for PC2. Spatial

autocorrelation = 0 at c. 135 meters.
Fig 3.3.15 Spatial correlogram for Betweenness_car.

Spatial autocorrelation = 0 at c. 125 meters.
78
Fig 3.3.16 Spatial correlogram for Betweenness_ped. Spatial autocorrelation = 0 at c.
152 meters.
Table 3.3.17 Summary of how far positive spatial autocorrelation extends for different variables.
79
3.3.2. Spatially constrained clustering
Fig 3.3.18 shows spatially constrained k=2 clustering based on variables PC3-4 plus location variables.
The PC variables are chosen because clustering based on them performed best in our comparison of
clusterings in section 2.3.3. The location variables added to this clustering are Latitude and Longitude.
The mean subjective scores in the spatial clusters have a slightly smaller difference in means than
those in the non-spatially constrained clusters (Fig. 3.3.18-c).
However, a colocation map is again the best way to evaluate the clustering. This is shown in Fig
3.3.19, which compares outcomes for spatially constrained and unconstrained clustering. Spatially
constrained clustering aligns better with subjective scores, giving an accuracy of 66.7% vs 59.6%
for regular clustering based on PC3-4. Note that in this case, cluster names were flipped so that the
cluster with the lower average subjective score was renamed “1” and the cluster with the higher
score was renamed “2”. After renaming, 33 data points in cluster 1 have a subjective score of 1, while
109 points in cluster 2 have a subjective score of 2. The clustering has particular trouble classifying
data points in the North and along Zurlindenstrasse in the eastern part of the route.
Fig 3.3.20 presents a similar analysis for objective scores. This time, non-spatial variables used are
PC1-2 because clustering on them perfumed best in section 2.3.3. This time, adding spatial variables
to the clustering does not increase standard deviation between cluster means compared to non-
spatial clustering. Additionally, comparing the clustering map with the map of objective scores in
section 3.3.1. does not show that clusters correspond closely to particular scores. Due to the spatially
scattered nature of objective scores, it would be very difficult to capture points with the same
objective scores within clusters which aim to maximise spatial contiguity.
Fig. 3.3.18 a) Spatially constrained clustering with k=2 based on PC3-4 data, longitude and latitude. b) The ratio of
between to total sum of squares is 0.23. c) Frequency of subjective values in each cluster.
80
a)
Accuracy: 142/213 = 66,7%
b)
Accuracy: 127/213 = 59,6%
Fig. 3.3.19 Colocation map for clustering results and subjective scores when a) spatial variables are included in clustering
and b) when clustering is based only on on PC3-4 and no spatial variables are included. Cluster names were flipped so
that cluster 2 is now the cluster with more subjective score = 2 values and vice versa. This map shows how often the data
points both belong to cluster 2 and have a subjective score of 2, or belong to cluster 1 and have a subjective score of 1.
Gray data points belong to cluster 1 and have subjective score 2 or vice versa.
81
a)
b)
c)
Fig. 3.3.20 a) K-means clustering with k=4 for PC1-2 and latitude and longitude. b) The ratio of between to total sum of
squares is 0.42. c) Distribution of objective score values in the clusters.
82
3.3.3. Qualitative analysis
Following the same procedure as the Weimar analysis, Fig. 3.3.21 summarises interesting locations
for different variables in the Zürich dataset. The figure is obtained by analysing break points in
the choropleth maps in section 3.3.1. Fig 3.3.22 puts these breakpoints in the context of a textual
description of the video. No clear trends are found besides the fact that break points are likely to be
located at intersections. In one case, the cause for a break point appears to be stairs leading from
a regular street to a smaller, quieter one below it (see interval 9:45-10:35, or data points 117-127).
Perhaps busy roads can also cause break points: sudden bouts of traffic or a group of people leaving
the bus simultaneously can quickly change environmental conditions. Quieter streets probably tend
to be more stable.
Fig. 3.3.21. A summary break points for different variables. The squares highlight places where a variable’s values undergo
an important shift, such as suddenly going from a very low to a very high value. Places where values change very frequent-
ly were not marked, however. The background map shows subjective scores.
83
Fig 3.3.23 contextualises subjective scores. Positive subjective scores (score =1) are focused on the
northern and southeastern parts of the region, and like in Weimar, these regions tend to contain
more mentions of trees than other regions. These regions also have more mentions of pedestrians;
however, the significance of the finding is questionable since there are generally so few pedestrians
in the area that even a few passerbys are an exception. Perhaps more notably, there are no mentions
of heavy traffic or fast cars in areas with positive ratings, implying that these features may be related
to positive perceptions. While watching the video, I personally felt averse to places with high amounts
of traffic noise and cars zooming past. It is difficult to draw many other conclusions from the video. In
my opinion, the environment in Zürich did not contain significant high or low points.
Since negative comments about traffic in the video notes suggests that there is a link between
environmental attractiveness and number of cars, this topic is studied further with the help of Fig
3.3.24. Betweenness_car values and video notes do not align very well. One area with “slightly
heavier traffic”, according to the notes, belongs to the bottom quartile of Betweenness_car. A region
in the North described as having “light traffic” belongs to the top quartile of Betweenness_car.
Perhaps this shows that human perceptions of disruptive traffic do not depend only on the number
of cars but also how close those cars are and how fast they drive. Watching the video again, I
notice that there are indeed more cars in the places shown as having high Betweenness_car values.
However, when watching the video the first time I was more bothered by loud and fast cars in a busy
intersection than slow cars next to a park.
Fig. 3.2.25 combines notes on the video with a map of Betweenness_ped values. Watching the video,
it appeared there were almost no pedestrians anywhere. The only clear pedestrian hotspot occurred
at the three-way intersection in the middle of the region after a bus unloaded a group of passengers
there. In addition, sometimes several pedestrians gathered at intersections to wait for their turn to
cross the street. The southern street with the top quantile Betweenness_ped values is not identified
in the video notes as a pedestrian hot spot. These results suggest that Betweenness_ped values
should not be given too much consideration in the Zürich data. Pedestrian numbers are so low that
random occurrences like bus stops and slow traffic lights have significant influence on whether a
place is identified as a pedestrian hot spot. Perhaps the measure is significant for long streets where
there are no pedestrians at all, such as in the video at 1:05 to c. 3:00 (Fig. 2.3.6). Zero pedestrians can
be a sign of an unattractive or boring street. After watching the video, however, it is clear that high
numbers of pedestrians don’t tell us much about environmental quality in this dataset.
Figs 2.3.26-7 present a sample of screen shots from places with consistently neutral and positive
subjective scores respectively. Neutral images are taken from the “boring” street with very few
pedestrians described above (video 01:00-03:00). Positive images are taken from the northeastern
part of the area (video c. 04:35-08:00). Neither set of images features any pedestrians. The feature
which most clearly distinguishes the group with positive ratings is the abundance of trees and plants
there. The pictures are also more varied among each other and contain more open space. The images
from the street with neutral ratings have more subdued colors and conventional buildings, and don’t
offer obvious places to relax or stop to enjoy oneself.
84
A neighborhood park with Intersection with a few An open space with trees,
many trees, a café with trees visible further ahead. buildings visible, a few
seating outside (though A shopfront. Sidewalk next cyclists and pedestrians.
no people). Walking on a to a very long residential Walking towards the open
sidewalk next to a road building ‘s hedge bush on space on a sidewalk beside
with light traffic. 1 one side and parked cars trees and grass on one
Shop with a tree in front of it. on the other. (05:45-
pedestrian and 1 cyclist. side and a car road on the
Person unloading a truck. Some 06:50, 69-82)
(04:35-05:40, 53-68) other. (06:55-07:45, 69-
trees visible ahead. Buildings
93)
unremarkable. (04:05-04:50, 49-
54)
Similar house with a
hedge bush as before.
Quieter residential streets with empty parking spaces
little traffic, many cars parked. No next to sidewalk.
people, trees or shop windows.
Lots of concrete. 4-storey ordinary
Busy intersection with
residential houses. Very few
cars driving fast. A
people (3 peds and a cyclist during
few pedestrians
the entire interval) (01:05-04:00,
12-48)
Residential street with
some walled gardens.
Slightly heavier traffic, shop 1 car passes quickly.
windows, no trees. 1 pedestrian. Lots of trees. A few
Cars parked. (00:40-01:00, 8-11) pedestrians. Sound of
traffic.
Almost no people. 5-storey
residential buildings, quite modern. Quiet residential
A few trees and shop windows. street with a narrow
Light traffic. (00-00:35, 0-7) sidewalk and road.
Hedge next to
Bus load of sidewalk. No people.
Many shops, hair people cars parked.
dressers etc. Some
buildings have Café with seats
private gardens with outside but no
trees. Standard people. Just a few
apartment buildings. people standing
No people. around. No trees,
cars next to sidewalk.
Bus stop, trees Approaching
at intersection. intersection. No p
Lots of concrete, no trees or
Historical- trees, or people,
people. Show window. Some
looking Modern many streets open
Historical graffiti on buildings. Passing
apartments. buildings. Shops. in different
looking under a portico of a nice stone
Bikes stacked on directions.
apartments. building.
Almost no side of street. A
trees except few very small
for a few trees next to
private sidewalk. 1
gardens. No pedestrian.
people.
Fig. 3.3.22. VariableQuite many in the context of the textual description of the video. As in Weimar, the breakpoints 91
breakpoints
often concur with textual description breakpoints. Environmental features mentioned in the text are highlighted using
cars passing,
the same color as indrive
the break
prettypoint boxes. An exception to this are trees and shops (represented by green and red
respectively), whichfast.
were not included in the raw data and which therefore do not have any break points.
85
p
Fig. 3.3.23. Subjective scores in the context of the textual description

of the video. Places with positive subjective scores in the North-West
seem to have more greenery than other places. The other variables
highlighted in the description do not seem correlated with scores.
86
p
Fig. 3.3.24. Betweenness_car in the context of the textual description

of the video. High betweenness_car values and textual descriptions of
heavy traffic do not always correlate.
87
p
Fig. 3.3.25. Betweenness_ped in the context of the textual

description of the video. High betweenness_ped values and
textual descriptions of many pedestrians do not correlate
much. There are very few pedestrians in the area overall.
88
Fig. 3.3.26 Screenshots from the video at 1:00, 2:00 and 3:00 minutes (from the top). These
images are from the quiet neighborhood street in the South-West which consistently received
subjective scores of 2.
89
Fig. 3.3.27 Screenshots from the video c. 4:45, 5:20, 6:40, 7:45 (from the
top). These images are from the North-Eastern side of the experiment’s
area, which consistently received subjective scores of 1.
90
3.3.4. Extended qualitative analysis
For the Zürich data, it is possible to expand the qualitative analysis conducted above. This is because
a previous study on Zürich is available which precisely maps different features of the environment,
including, for example, traffic speeds, trees, and building construction dates (Schaeben, 2017). These
maps can be analysed together with subjective scores to make our qualitative analysis more precise.
Figs 3.3.29-31 overlay subjective scores with maps of greenery, buffer typology and facade
transparency created by Schaeben (2017). Subjective scores are represented by lines which simplify
the original subjective score data points (Fig. 3.3.28). Since the buffer types include “semi-private
green” and “private green”, greenery and buffer typology are related, and both appear to be
correlated with subjective scores. More greenery always occurs in places with positive subjective
scores, with the exception of Zurlindenstrasse in the East (Fig. 3.3.29). Meanwhile places with neutral
subjective scores have quite an even mix of different types of buffer, though parking is slightly
dominant. It therefore appears that greenery is more correlated with subjective scores than buffers,
and it is primarily the green nature of certain buffers that causes a correlation with subjective scores.
Fig 3.3.32 overlays subjective scores with a traffic speed map. It appears to confirm our
previous hypothesis that traffic speed is an important factor in human perceptions of the urban
environment. Positive subjective scores correlate almost completely with one-way streets with 30
km/h speed limits. Traffic may also explain why Zurlindenstrasse gets consistently neutral reviews
despite its abundance of trees: after the one-way street in the North, pedestrians may feel that
Zurlindenstrasse’s two-lane traffic represents a noticeable increase in traffic. This is especially true
since walking along Zurlindenstrasse requires one to cross a road with a 50 km/h speed limit.
Finally, Schaeben’s maps enable us to briefly consider the impact of buildings’ architectural style
on human perceptions - something which is not quantified in our current data set nor discussed
extensively in the textual summary of the videos. Fig. 3.3.33 overlays subjective scores with a map
of the area’s buildings’ construction dates. Almost all positive subjective scores are surrounded by
buildings built before 1960, with the oldest buildings dating from the late 19th century. However, the
area only has about 10 buildings which are built after 1960 and which are located directly beside the
experiment’s path. Given the small sample size, it is not possible to form any far-reaching conclusions
about which architectural style people prefer.
91
Fig. 3.3.28. Simplified subjective scores. The green and Fig. 3.3.29. Simplified subjective scores on a greenery map.
blue lines represent broader areas of subjective score Except for the more southern park area, trees and hedges
1 (positive) and 2 (neutral), while the dots represent seem to correlate with positive subjective scores. The
original data points. Lines were not drawn in places greenery base map is by Schaeben (2017).
where subjective scores change frequently. The simplified
lines are used in the following maps.
Fig. 3.3.30. Simplified subjective scores on a buffer Fig. 3.3.31. Simplified subjective scores on a facade
typology map. Except for the parking in the north and the transparency map. Facade transparency does not seem to
southern park area, semi-private green buffers and gardens be strongly correlated with subjective scores. Areas with
appear to be most clearly correlated with positive subjec- positive scores have many apartment facades, but on the
tive scores. The buffer base map is by Schaeben (2017). other hand, other areas do too. The facade base map is
created by Schaeben (2017).
92
Fig. 3.3.32. Subjective scores overlaid on a traffic speed map. Almost all positive scores are located on one-way streets
with a 30 km/h speed limit. On the other hand, the majority of neutral scores are located on streets with two-way traffic
and/or greater speed limits. The traffic speed base map is by Schaeben (2017).
93
Fig. 3.3.33. Subjective scores overlaid on a construction date map. Positive scores are mostly located beside buildings built
before the 1960s. However, there are also many neutral scores in the South-West surrounded by buildings built in the
same period. Construction dates therefore do not seem to be strongly correlated with subjective scores. The traffic speed
base map is by Schaeben (2017).
94
3.4. Summary of spatial analysis results
The spatial analysis brings several interesting conclusions. The study finds, firstly, that spatial
autocorrelation is significant for both human perceptions and raw data variables. Secondly, spatially
constrained clustering may improve clustering results and the associated colocation map certainly
provides useful insights into clustering performance. Thirdly, a combined qualitative and spatial
analysis can encourage critical thinking of data analysis results and help to find new features which
explain human perceptions, most notably greenery and traffic speed.
The spatial autocorrelation analysis in sections 3.2.1. and 3.3.1. shows that significant spatial
autocorrelation exists for almost all variables in both Zürich and Weimar data. This can be easily
seen by examining choropleth maps and quantified more precisely with Moran’s I. This study also
investigates the extent of spatial autocorrelation for different variables, finding that most variables are
not only spatially autocorrelated to their neighbours, but also to data points up to 200 meters away.
This information is not necessarily useful for further analyses, but it does strengthen the finding of
spatial autocorrelation for all variables.
Spatial autocorrelation may also help to improve the performance of clustering. This is suggested by
the comparison between spatial and ordinary k=2 clustering on Weimar and Zürich data. For Weimar,
its spatial clustering accuracy is 77.7% vs 61.3% for ordinary clustering. For Zürich, its accuracy
is 66.7% compared to 59.6% for ordinary clustering. The improvements probably occur because
subjective scores are highly spatially autocorrelated. For Weimar especially, scores quite neatly form
two clusters. In these conditions, it helps to have a clustering method which seeks to create spatially
contiguous clusters rather than spatially separated ones.
Colocation maps can help to understand how the clustering could be improved further in the future.
The colocation map offers similar data as a confusion matrix (how many data points ended up in the
correct cluster), however its spatial dimension brings extra benefits. With the map, it is possible to
identify the broad regions where clustering performed well or badly. In the case of Weimar data,
clustering failed in specific parts of the region. Based on this knowledge, one could attempt to identify
the confounding factors in those areas and then incorporate these factors in future clustering. This is
easier to do for a specific region than for a collection of individual points whose links to one another
are not understood.
The combined spatial and qualitative studies in sections 3.2.3 and 3.3.3. take a step in this direction.
By combining choropleth maps of for example subjective scores with a textual description of
the environment, the study attempts to improve our understanding of why certain regions
have consistently high or low subjective scores. Features such as perceived amount of trees,
pedestrians, shops and cars were considered. For Weimar, trees, pedestrian areas and pedestrian
congestion seemed correlated with high subjective ratings. For Zürich, only trees appeared to have a
straightforward correlation. Importantly, however, the Zürich study helps to answer the question of
why high numbers of pedestrians don’t correlate with high subjective scores. Through the qualitative
analysis it is found that in Zürich, high numbers of Betweenness_ped scores don’t actually reflect lively
streets, but rather chance occurrences such as a group of people getting off a bus. This is a valuable
insight which could not have been reached through processing quantitative data alone.
The qualitative analysis in Zürich also brought other insights. The textual video analysis already
suggested that Betweenness_car may not always be the most appropriate metric for measuring
how much people are affected by cars, and that factors like speed, noise and distance from the car
may also play a role. This finding was reinforced by the combined traffic speed - subjective score
map. A similar map based on Schaeben’s work also strengthened the link between abundance of
greenery and positive subjective scores. Such maps seem like a good approach to understanding
95
human perceptions of the built environment. Once done, they can also provide the basis for further
quantitative research. By simply giving the different greenery or speed categories numerical values,
they could be incorporated into our raw data, so that their links with perceptions could be rigorously
explored with the analytical techniques of the non-spatial analysis.
96
4. CONCLUSIONS
This section draws together the conclusions of non-spatial and spatial data analysis. The research
did not uncover any novel relationships between human perceptions and features of the built
environment. However, the principles and methodology of the study provide a promising starting
point for future work.
In the non-spatial data analysis, regression, PCA, k-means clustering and decision trees are used to
investigate the relationship between environmental features and human perceptions. Objective and
subjective human perceptions are not correlated with each other and subjective scores are found to
be much more reliably linked to spatial features. The variable which is most clearly linked to human
perceptions is Betweenness_ped, while RayPortion obstacles, Betweenness_ car and Min_radial
might be important too. Despite its importance, Betweenness_ped is also contradictory. Although
extensive literature exists on the positive correlation between urban attractiveness and the number
of people in the environment, Betweenness_ped correlates negatively with Zürich subjective scores.
Decision trees trained on Zürich data are also very unsuccessful at predicting subjective scores in
Weimar, even though Betweenness_ped is important in each and is a supposedly universally positive
attribute. The conclusions of the non-spatial analysis are therefore slightly unsatisfying: subjective and
objective scores are uncorrelated; most variables do not show a strong relationship with either one;
and Betweenness_ped, the most promising candidate for such a correlation, has very different kinds
of relationships to subjective scores in Zürich and Weimar data.
The spatial analysis succeeds in clarifying some of these issues. First, by showing that spatial
autocorrelation exists for all variables, the spatial analysis demonstrates that x-variables and
subjective/objective scores are not distributed randomly in space. This is important because it means
that inconclusive results are not caused by randomness of people’s perceptions. Rather, it suggests
that correlations between physical features and perceptions can be uncovered if we improve our
metrics for describing the physical features of space.
The spatial analysis also sheds light on the mystery of Betweenness_ped’s relationship with subjective
scores. Combining choropleth maps of Betweenness_ped with textual descriptions of the Weimar
and Zürich videos shows that Zürich has much fewer pedestrians overall. This information can, of
course, be gained by simply investigating Betweenness_ped ranges for both datasets. The unique
contribution of the qualitative analysis is an explanation that in Zürich, random events such as a
bus dropping people off can create heightened Betweenness_ped values which have no relation to
subjective scores. Weimar, with its greater baseline amount of pedestrians, is influenced less by these
kinds of disturbances. Based on this explanation, we can conclude that Betweenness_ped is indeed
positively correlated with subjective scores, however insufficient sample size in Zürich prevents this
relationship from manifesting.
In addition, the combined spatial and qualitative analysis suggests that the feature Betweenness_car
may not be the most important car-related attribute which influences human perception of the built
environment. When viewing the videos, my reaction towards fast, close and noisy cars was much
more negative than towards cars which drove slowly, far away or didn’t make much noise. A further
analysis of traffic speed limits’ alignment with subjective scores supports this conclusion. Perhaps
if speed and noise were incorporated into the data, cars would have a bigger impact on people’s
subjective and objective evaluations of environments. Another feature which the qualitative analysis
97
suggest would be important to include in the data is the amount of trees or vegetation. For both
Weimar and Zürich, there appeared to be a clear positive correlation between amount of vegetation
and subjective ratings.
This study’s findings are also relevant for architectural research. As mentioned in the Introduction,
there is a current of architectural research which aims to utilise new technologies to gather data and
analyse people’s relationship with their surroundings. Because it examines the same isovist variables
as for example Li (2016) and Hijazi (2016), our study’s findings can be directly compared with theirs.
Notably, unlike these previous studies, our analysis did not find isovist variables such as compactness,
perimeter or occlusivity to be significantly correlated with human perceptions. Instead, we would
recommend incorporating new variables such as greenery and traffic speed into future studies. More
fundamentally, our research questions the reliability of physiological metrics in communicating how
positively a person perceives their surroundings, since we found very weak or no correlation between
these and people’s conscious evaluations of the urban environment.
The methods used in this study could also benefit architectural researchers and practitioners. Spatial
analysis methods, in particular, seem like an under-utilised asset for architects and urban planners.
Spatially constrained clustering, colocation maps and even simple choropleth maps could be very
useful for urban planners attempting to understand and visualise features of the urban area such as
the spatial distribution of different-age people in the neighborhood. In the long run, urban planning
and research will probably shift from an intuition-based towards a data- and model-based direction
as computational methods begin to outperform humans in increasing numbers of analytical tasks.
Such data-based approaches could improve the quality and efficiency of design-work, and I therefore
believe that researching the use of these methods in architecture is worthwhile.
98
5. BIBLIOGRAPHY
Anselin, L., (2017). GeoDa: An Introduction to Spatial Data Analysis. Retrieved from: http://geodacen-
ter.github.io/documentation.html
Bus, P. (2016) Visibility in Space Syntax [lecture slides]. Retrieved from: http://www.ia.arch.ethz.ch/
wp-content/uploads/2016/10/lecture_03.pdf
Clifton, K., Ewing, R., Knaap, G. J., & Song, Y. (2008). Quantitative analysis of urban form: A multidisci-
plinary review. Journal of Urbanism, 1(1), 17-45. DOI: 10.1080/17549170801903496
ESUM, (2017), Human perception of urban environment, https://www.esum.arch.ethz.ch
Gehl, J., (1979/2011). Life Between Buildings: using public space. Washington DC: Island Press cop.
Girardin, F., Vaccari, A., Gerber, A., Biderman, A., Ratti, C., (2009). Quantifying urban attractiveness
from the distribution and density of digital footprints. International Journal of Spatial Data
Infrastructures Research, 4, 175-200. Retrieved from: https://pdfs.semanticscholar.org/0f52/337ec-
7be543cd046060ed8ef1de8e83ed07e.pdf?_ga=2.42858220.851510460.1536628711-
1868991344.1536392739
Hijazi, I., Koenig, R., Schneider, S., Li, X., (2016). Geostatistical Analysis for the Study of Relationships
between Emotional Responses of Urban Walkers to Urban Spaces. International Journal of E-Planning
Research, 5(1). Retrieved from: https://www.researchgate.net/publication/304272875/download
Jacobs, J.,(1961/1993). The Death and Life of Great American Cities. New York: Modern Library
Kent State University. Cluster Validation. Retrieved from: http://www.cs.kent.edu/~jin/DM08/Cluster-

Validation.pdf
Li, X., Hijazi I., Koenig, R., Lv, Z., Zhong, C., Schmitt, G., (2016). Assessing Essential Qualities of Urban
Space with Emotional and Visual Data Based on GIS Technique. International Journal of Geo-Informa-
tion, 5(11), 218. doi: 10.3390/ijgi5110218
Mehta, V., (2013). Evaluating Public Space. Journal of Urban Design, 19(1), 53-88. doi:
10.1080/13574809.2013.854698
Resch, B., Sudmanns, M., Sagl, G., Summa, A., Zeile, P., Exner, J., (2015). Crowdsourcing Physiological
Conditions and Subjective Emotions by Coupling Technical and Human Mobile Sensors. GI Forum -
Journal for Geographic Information Science, 2015(1), 514. doi: 10.1553/giscience2015s514
Schaeben, C. (2017), People’s Perception of Urban and Architectural Features. Retrieved from:
https://www.research-collection.ethz.ch/handle/20.500.11850/266634
Sauwen, N., Functions in NMF. Retrieved from: https://www.rdocumentation.org/packages/NMF/

versions/0.21.0/topics/purity
United Nations, (2018). World Urbanization Prospects: the 2018 Revision: Key Facts. Retrieved from:
https://population.un.org/wup/Publications/Files/WUP2018-KeyFacts.pdf
99

Thesis Heidi Silvennoinen

Uploaded by

Copyright:

Available Formats

Thesis Heidi Silvennoinen

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thesis Heidi Silvennoinen

Uploaded by

Copyright:

Available Formats

Research Collection

Non-Spatial and Spatial Statistics for Analysing Human

Advisor: Dr. Varun Ojha

2. Non-spatial data analysis

2.1. Summary of method..................................................................................................................................................... 12

3. Spatial data analysis

3.1. Summary of method..................................................................................................................................................... 49

Fig. 1.1. Devices used in the experiment............................................................................................................. 8

2.2. Weimar non-spatial data analysis

Fig. 2.2.1. Correlation between objective and subjective scores........................................................................... 14

2.3. Zürich non-spatial data analysis

Fig. 2.3.1. Correlation between objective and subjective scores ........................................................................ 30

3.2. Weimar spatial data analysis

Fig. 3.2.1. Chroropleth maps of subjective and objective scores........................................................................... 50

3.3. Zürich spatial data analysis

Fig. 3.3.1. Chroropleth maps of subjective and objective scores........................................................................... 72

2.2. Weimar non-spatial data analysis

2.3. Zürich non-spatial data analysis

2.4. Summary of results of non-spatial data analysis

3.2. Weimar spatial data analysis

3.3. Zürich spatial data analysis

1.3. State of the Art

1.4. Virtual Reality Experiment

Fig. 1.1. a) 14 GoPro Hero 3+ cameras

Table 1.2. Explanation of the ten environmental variables in our data.

The research questions in this study are:

2.1. Summary of method

2.1.1. Data Preprocessing

The non-spatial data analysis consists of the following parts:

3. K-means clustering on to check whether clusters correspond to objective / subjective ratings:

2.2.1. Regression analysis

a) Betweenness_ped vs Subjective score. r = -0.38, p = 3.2*10 - 9

Fig. 2.2.11. Principal components 1-5 plotted

Fig. 2.2.12. Principal components 1-5 plotted

Mean of all data

Fig 2.2.14 Betweenness_ped vs

Mean of all data

2.2.4. Feature extraction using decision tree

2.3.1. Regression analysis

Table 2.3.8. Table showing contributions (in %)

Fig 2.3.9. Contributions of variables to different

Fig. 2.3.11. Principal components 1-4

2.3.4. Feature extraction using decision trees

Weimar, Weimar, Zürich, Zürich, Table. 2.4.1. Summary

Table. 2.4.2. Summary of

Min_radial in each tree. The green part states

own data set's

3.1. Summary of method

1. Spatial autocorrelation: quantifying the influence of location on attributes

2. Spatially constrained k-means clustering

3.2.1. Spatial autocorrelation

Setting spatial weights

Fig 3.2.6 Spatial autocorrelation for objective scores.

Fig 3.2.7 Spatial autocorrelation for subjective scores.

Fig 3.2.8 Spatial autocorrelation for PC1.

Fig 3.2.10 Spatial autocorrelation for Betweenness_car.

Fig 3.2.11 Spatial autocorrelation for Betweenness_ped.

Fig 3.2.13 Spatial autocorrelation for RayPortion_obstacles.

Fig 3.2.14 Spatial correlogram for subjective scores.