Remote Sensing of Environment 183 (2016) 170–185

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Detailed dynamic land cover mapping of Chile: Accuracy improvement

by integrating multi-temporal data
Yuanyuan Zhao a, Duole Feng a, Le Yu a, Xiaoyi Wang b,c, Yanlei Chen c, Yuqi Bai a, H. Jaime Hernández d,
Mauricio Galleguillos e,h, Cristian Estades f, Gregory S. Biging c, John D. Radke g, Peng Gong a,b,c,⁎
a
Ministry of Education Key Laboratory for Earth System Modelling, Centre for Earth System Science, Tsinghua University, Beijing 100084, China
b
State Key Laboratory of Remote Sensing Science, Jointly Sponsored by Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, and Beijing Normal University, Beijing 100101,
China
c
Department of Environmental Science, Policy, and Management, UC Berkeley, Berkeley, CA 94720, USA
d
Geomatics and Landscape Ecology Lab, Forestry and Nature Conservation Faculty, University of Chile, Santiago, Chile
e
Department of Environmental Sciences, Faculty of Agronomic Sciences, University of Chile, Santiago, Chile
f
Wildlife Ecology Lab, Forestry and Nature Conservation Faculty, University of Chile, Santiago, Chile
g
Department of Landscape Architecture and Environmental Planning, UC Berkeley, Berkeley, CA 94720, USA
h
Center for Climate and Resilience Research (CR)2, University of Chile, Santiago, Chile

a r t i c l e i n f o a b s t r a c t

Article history: Stretching over 4300 km north to south, Chile is a special country with complicated landscapes and rich biodiver-
Received 22 August 2015 sity. Accurate and timely updated land cover map of Chile in detailed classification categories is highly demanded
Received in revised form 11 May 2016 for many applications. A conclusive land cover map integrated from multi-seasonal mapping results and a sea-
Accepted 23 May 2016
sonal dynamic map series were produced using Landsat 8 imagery mainly acquired in 2013 and 2014, supple-
Available online 6 June 2016
mented by MODIS Enhanced Vegetation Index data, high resolution imagery on Google Earth, and Shuttle
Keywords:
Radar Topography Mission DEM data. The overall accuracy is 80% for the integrated map at level 1 and 73% for
Land cover mapping level 2 based on independent validation data. Accuracies for seasonal map series were also assessed, which is
Landsat around 70% for each season, greatly improved by integrated use of seasonal information. The importance of
Seasonal dynamics growing season imagery was proved in our analysis. The analysis of the spatial variation of accuracies among var-
Biodiversity ious ecoregions indicates that the accuracy for land cover mapping decreases gradually from central Chile to both
30 m north and south. More mapping efforts for those ecoregions are needed. In addition, the training dataset includes
sample points spatially distributed in the whole country, temporally distributed throughout the year, and cate-
gorically encompassing all land cover types. This training dataset constitutes a universal sample set allowing
us to map land cover from any Landsat 8 image acquired in Chile without additional ad hoc training sample
collection.
© 2016 Elsevier Inc. All rights reserved.

1. Introduction spatial scales, the quality of land cover maps are hardly satisfactory to
meet the needs of diverse user communities (Yu et al., 2014a). A large
Land cover information plays a key role in many Earth system stud- gap in land cover mapping is the shortage of land cover maps at the na-
ies (Yang et al., 2013). It is an important input to Earth system models tional scale for many countries in the world. Global land cover maps
(Dai et al., 2003; Jung, Henkel, Herold, & Churkina, 2006; Liu, Chen, cannot fully fill this gap because their land cover categories are usually
Cihlar, & Park, 1997) or habitat models (Liang et al., 2010; Özesmi & designed for specific global applications. Given the considerably im-
Mitsch, 1997; Pearson, Dawson, & Liu, 2004; Yu, Shi, & Gong, 2015). It proved availability of medium resolution satellite data and data pro-
is also essential for natural resources planning and management cessing capabilities, a critical task in producing useful land cover maps
(Gong, 2012; LaFontaine, Hay, Viger, Regan, & Markstrom, 2015; for a particular country is adopting a land cover classification scheme
Pauleit & Duhme, 2000; Zhong, Gong, & Biging, 2012). Although a developed with the participation of users in that country. Chile is such
large number of efforts have been made to map land cover at various a country that does not have a comprehensive land cover map at medi-
um spatial resolution. This study describes our national land cover map-
ping effort for Chile.
⁎ Corresponding author at: Ministry of Education Key Laboratory for Earth System
Bounded by the Pacific on the west, the Andes on the east and the
Modelling, Centre for Earth System Science, Tsinghua University, Beijing 100084, China. Atacama Desert in the north, Chile is a virtual continent island
E-mail address: [email protected] (P. Gong). stretching over 4300 km north to south with rich biodiversity. Notably

http://dx.doi.org/10.1016/j.rse.2016.05.016
0034-4257/© 2016 Elsevier Inc. All rights reserved.
 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185 171

Table 1
Regional land cover products covering the entire Chile.

Product name Data time Resolution Sensor Scheme Reference

UMD SA 1987–1983 1 km AVHRR 16 classes Townshend et al. (1987)

LBA-ECO 1987–1991 1 km AVHRR 39 classes Stone et al. (1994)
JRC SA 1995–2000 1 km ATSR, VGT, OLS 10 classes (L1) Eva et al. (2004)
SERENA 2008 500 m MODIS 22 classes Blanco et al. (2013)
MERISAM2009 2009–2010 300 m MERIS 9 classes Hojas Gascon et al. (2012)
USGS SA 2010 30 m TM, ETM+ 5 classes Giri & Long (2014)

a high rate (about 45%) of the species are endemic, which can be attrib- sufficient spatial resolution to capture important phases in the vegeta-
uted to its isolation (Squeo et al., 2012). The climate regions vary from tion dynamics of the landscape. For example, a series of seasonal dy-
tropical desert and semi-desert in the north, Mediterranean in the mid- namic land cover map of high resolution could even capture the
dle, temperate oceanic and sub-polar oceanic in the south. Under the ef- moment when geophytes and annuals sprout from the seemingly bare
fects of both climate and topography, the natural vegetation types are land.
highly diverse. In the north lies the driest Atacama Desert, with almost Since the 1950s, the land cover type of southern-central Chile is
no vegetation at its heart, low scrub vegetation towards the Andes dominated by fast-growing commercial plantations, mainly including
and sparse shrub called the “Lomas” towards the coast, highly depend- radiate pine and eucalyptus (Toro & Gessel, 1999). Forests change main-
ing on coastal fog and humidity. After a transitional zone of matorral ly occurs in this region, with a loss of more than 1 million hectares and a
(shrubs) and savannas southward, the vegetation changes to Mediter- concurrent gain of about the same area from 2001 to 2013, according to
ranean sclerophyll woodlands and high shrubs. With the remarkable Global Forest Watch (an online platform for the public to help monitor
presence of the typical Araucanía forests and krummholz, the deciduous forest change, http://www.globalforestwatch.org/), reflecting the rapid
forests in central Chile transit towards temperate rain forests with high rotations of logging and replantation of trees. With the log exports being
precipitation levels. For the fjords and islands in the far south, it be- reduced, harvested wood is processed into primary and secondary
comes moorlands or icefields and oceanic forests reappear (Moreira- products, including paper, pulp, sawnwood, engineered wood products
Muñoz, 2011). Chile's complicated landscape and diverse vegetation re- and furniture, and 75% of the product is exported mainly to the United
quires a monitoring approach that uses multi-temporal imagery of States, Japan and China (Cartwright & Gaston, 2002). The carbon

Fig. 1. Flowchart of the Chilean land cover mapping process.

172 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185

Table 2 latest product for South America was 30 m resolution derived by USGS
Land cover classes in the classification scheme used in this study. for 2010 using mainly Landsat 5 TM and Landsat 7 ETM+ images (Giri &
Level 1 Level 2 Level 3 Long, 2014). They only used the Landsat images of growing season, but
100 croplands 110 rice fields
we believe there is more information we can extract from the images
120 greenhouse farming outside the growing season. They mapped 5 classes, including trees,
130 other croplands open water, barren, perennial snow/ice and other vegetation, but the
140 orchards classification scheme lacks details for many thematic applications. The
150 bare croplands/fallow
5 classes were conclusive land cover classes, but for many applications
200 forests 210 natural broadleaf 211 primary
212 secondary we need seasonal dynamics of the land cover classes. For validation,
220 natural conifer 221 primary they used 55 of the 500 validation sites (Olofsson et al., 2012) located
222 secondary in South America, and there are only 4 sites in Chile, and 3 of them are
230 natural mixed distributed in Central Chile, so the Chile part of the product has not
240 broadleaf plantations 241 old
plantations
been fully validated. There are also a few local maps focusing on specific
242 clearcuts scope of applications, for example, mapping a part of Araucanía Region
250 conifer plantations 251 old to study native forest loss and turnover (Altamirano et al., 2013), map-
plantations ping a part of Maule and Bío-Bío regions to study plantation change
252 clearcuts
(Nahuelhual, Carmona, Lara, Echeverra, & González, 2012), and map-
300 grasslands 310 pastures
320 other grasslands ping an area in the Valparaíso, O'Higgins and Metropolitan regions to
330 withered grasslands study the transformation from Mediterranean vegetation to intensive
400 shrublands 410 shrublands cultural landscapes (Schulz, Cayuela, Echeverria, Salas, & Benayas,
420 shrubs and sparse trees mosaic 2010).
430 succulents
Temporal information has long been one of the most important in-
440 shrub plantations
450 withered shrublands formation source for medium resolution land cover mapping (Arvor,
500 wetlands 510 marshlands Jonathan, Meirelles, Dubreuil, & Durieux, 2011; Friedl et al., 2010;
520 mudflats Hansen, DeFries, Townshend, & Sohlberg, 2000; Tucker, Townshend,
530 other wetlands
& Goff, 1985). With better data availability of multi-temporal high res-
600 water bodies 610 lakes
620 reservoirs/ponds
olution images, the use of more than one image for a single year was
630 Rivers proved useful on improving mapping accuracy. The most effective ap-
640 ocean plication of multi-temporal images is the croplands mapping (Brisco &
800 impervious Brown, 1995; Lo, Scarpace, & Lillesand, 1986; Peña & Brenning, 2015;
surfaces
Zhong, Gong, & Biging, 2014). Besides, multi-temporal images were
900 barren lands 910 dry salt flats
920 sandy areas also useful in mapping grasslands (Guerschman, Paruelo, Bella,
930 bare exposed rocks 931 rocks Giallorenzi, & Pacin, 2003; Henebry, 1993), forests (Fagan et al.,
932 gravels 2015; Wolter, Mladenoff, Host, & Crow, 1995) and wetlands (Dong et
1000 snow and ice 1010 snow
al., 2014; Munro & Touron, 1997). Most of these studies chose two to
1020 ice
1200 clouds
six images of different time to combine in the feature set (Homer et
al., 2007; Senf, Leitão, Pflugmacher, van der Linden, & Hostert, 2015;
Schriever & Congalton, 1995; Zhu, Woodcock, Rogan, & Kellndorfer,
2012), but the problem is we cannot ensure the same number of
high quality images covering different path/row locations for large
sequestration dynamics and carbon sink potential of plantations are also scale land cover mapping. Some researchers extracted some metrics
of wide interest (Espinosa, Acuña, Cancino, Muñoz, & Perry, 2005), from long or short time series (for example, median and mean of
which would be greatly facilitated by a long-time series of high resolu- three earliest and latest observations ordered by date) and added
tion plantation maps. Another important problem related to plantation them to the feature set (Potapov et al., 2015). However, in this way
is the concerns about ecosystem services. The diverse native forests we can end up with same data structure for classification inputs, but
were replaced by single species, affecting the habitat and biodiversity, we may have information loss in this process. This study selected a
water regulation, and soil fertility. Having a map of the plantation and light-weight feature set, and tried our best to make full use of our
native forests makes it possible for researchers to study the balance of multi-seasonal training sample set, because improving training sample
economic benefits and ecosystem services (Panayotou & Ashton, is the most direct way to improve classification performance. A com-
1992). Since additional stress caused by plantations comes from insect mon model was built for the classification of every image we have,
outbreaks (Lanfranco & Dungey, 2001), risk and vulnerability analysis no matter where it was, when it was acquired or how many images
is necessary, which needs the pattern of plantation and native forests for each season we have for the same area. Additionally, the study
as a basic input. areas in most of the previous dynamic land cover mapping studies
Land cover mapping has long been focused in Europe, North Ameri- that dealt with both conclusive land cover types and seasonal land
ca, and even Asia, resulting in better products over many of these areas. cover types were limited to regional or smaller scales (Dronova,
Except for global land cover products, a few regional land cover prod- Gong, Wang, & Zhong, 2015; Hess, Melack, Novo, Barbosa, & Gastil,
ucts were developed for South America (summarized in Table 1). The 2003; Sun, Zhao, Gong, Ma, & Dai, 2014; Wang et al., 2015). This
first generation of specialized land cover product for South America study is the first attempt to derive a dynamic land cover product at na-
was 1 km resolution using AVHRR NDVI time series acquired in April tional scale to depict the seasonal land cover variation of Chile, a coun-
1982 to March 1983 (Townshend, Justice, & Kalb, 1987), followed by an- try stretching over 4300 km north to south with different climate
other 1 km resolution product generated by visual interpretation cover- zones and topographic conditions.
ing 1987 to 1991 (Stone, Schlesinger, Houghton, & Woodwell, 1994), In this paper, we described a detailed land cover database of Chile for
1 km resolution product mainly derived by Joint Research Centre of 2014, including a conclusive land cover map and a series of seasonal dy-
the European Commission for 2000 (Eva et al., 2004), 300 m resolution namic land cover maps. The land cover maps were validated using an in-
MERISAM2009 for 2009 (Hojas Gascon, Eva, Gobron, Simonetti, & Fritz, dependent sample dataset and the accuracies were reported. The spatial
2012) and 500 m resolution SERENA for 2008 (Blanco et al., 2013). The variation of accuracy among different ecoregions was analyzed as well.
 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185 173

Fig. 2. The spatial distribution of training (left) and validation (right) sample set.

2. Methods Topography Mission (SRTM) DEM was used as well, due to the complex
topography in Chile.
Fig. 1 illustrates an overview of the process in this mapping project,
including the data collection and pre-processing, classification scheme
and sampling design, training sample collection, classification, and ac- 2.2. Classification scheme design and sample collection
curacy assessment.
The classification scheme was designed with Chilean geographers
2.1. Data collection and pre-processing and biodiversity researchers, mainly based on the scheme used in the
FROM-GLC project (Gong et al., 2013). Some classes were refined to sev-
A total of 1605 images acquired by Landsat 8 Operational Land Imag- eral subclasses, driven by the application demands on biodiversity con-
er (OLI) for 2013 and 2014 were downloaded from the USGS archives. servation, resource planning and land surface modelling. This
Normally, there are several scenes for each season (i.e., spring, summer, classification scheme includes 10 classes at level 1, 30 level 2 and 35
autumn and winter in astronomic definition) at a certain Worldwide level 3 classes (see Table 2) which are potentially separable with avail-
Reference System 2 (WRS2) path/row location, but images with more able data. Most of the classes could be defined using the Land Cover
than 20% cloud cover were dropped from the mapping dataset to reduce Classification System (LCCS) method, which is a common language for
the computation workload, resulting in insufficient data for some spe- description and characterization of land cover features, for better com-
cific seasons over certain locations. To ensure at least one scene for parison and correlation of land cover classes in different classification
each season at each path/row location, three scenes with the least schemes (Di Gregorio, 2005). With additional information on canopy
cloud cover were used as supplements, including some newly acquired coverage and height, the classes in this scheme can be cross-walked bet-
in 2015. All these scenes were then radiometric calibrated and atmo- ter to others, such as FAO land cover classification system and the IGBP
spheric corrected using the Global Mapper (GM) software package de- land cover classification system. These detailed classes enable the land
veloped by our group (Gong et al., 2013). To support the sampling, cover maps to be used as fundamental data in applications of different
interpretation, and post-classification analysis, ancillary data includes fields. For example, to provide adequate information for forest manage-
the time series of a 16-day composite of Moderate Resolution Imaging ment and protection, plantation and natural forest were separated. The
Spectroradiometer (MODIS) Enhanced Vegetation Index (EVI) for the subclasses of leaf type can be used for land surface parameterization for
period of the years 2011 to 2014, a vector land cover database con- models. The differentiation of herbaceous croplands, orchard, and pas-
structed by Chilean Forestry Services, and high resolution imagery on ture made it possible to conduct further agricultural analysis, such as ag-
Google Earth (Zhao et al., 2014). 30 m resolution Shuttle Radar ricultural mode and its impact on biogeochemical cycles.
174 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185

Fig. 3. Multi-temporal Landsat image interpretation. Multi-temporal Landsat images were shown and the curves next to each window of Landsat image are the spectral profiles of the
center pixel. The horizontal axis represents the wavelength of each spectral band, and the vertical axis represents the reflectance. The supporting data of MODIS EVI observations and
high resolution image on Google Earth for the same location were also presented.

Fig. 4. Classification accuracies of the image acquired in each month by the model trained by training samples of each month, each season, or the whole year (feature set only includes
spectral bands of single image).
 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185 175

Fig. 5. Examples showing the differences of NDVI time series derived from multi-temporal Landsat images among land cover types including shrublands, grasslands, barren lands and
croplands. The green dots are NDVI observations of multi-temporal Landsat images, and the orange dots are MVC results for each season.

The training and validation datasets were sampled independently generalized polygons, labelled with major land cover classes. The spatial
using stratified random sampling based on the prior knowledge provid- distribution of training and validation sample set are presented in Fig. 2.
ed by the Chilean Forestry Services Database, which contains Since the training sample set should be relatively homogeneous,

Fig. 6. Examples showing the differences of NDVI time series derived from multi-temporal Landsat images among land cover types including natural forests and plantations. The green dots
are NDVI observations of multi-temporal Landsat images, and the orange dots are MVC results for each season.
176 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185

comprehensive and representative, 160 by 160 pixel image blocks of all accumulatively 4000 km of field trips were taken as far north as
phenological stages for all 2957 sampled “seed location” were extracted Atacama and as far south as Los Lagos. Approximately one thousand
for manual interpretation. One to five training sites of the land cover field points were carefully collected for validation.
type labelled for the polygon in Chilean Forestry Services Database
were selected from each image block and labelled with a time series 2.3. Classification strategy
of land cover class according to multi-temporal Landsat images. Fig. 3
demonstrates the multi-temporal interpretation of a training sample Random Forest was chosen in this study, because it is an efficient su-
in central Chile as an example. Dominated by subtropical high pressure pervised classifier with top-class performance among many machine
during summer, the precipitation is limited resulting in the seasonal al- learning methods (Gong et al., 2013). Random Forest is an ensemble
ternation of grassland and withered grassland. The curves next to each machine learning method that fits a multitude of decision trees using
window of Landsat image are the spectral profiles of the center pixel. subsets of training samples and integrates predictions of the individual
We can see the changing shape of the spectral profiles through multi- trees to improve classification accuracy and control over-fitting. The pa-
temporal observations. For example, from Dec. 31 to May 5, the spectral rameters for the classifier were determined from preliminary experi-
profiles present a wilting process of green vegetation by the decreasing ments. For example, we chose 100 as the number of trees in the
red absorption and near-infrared reflection, followed by a green-up pro- forest, which achieved a good balance between prediction performance
cess after May 5 showing an increasing red absorption and near-infra- and cost in computation time. With corresponding land cover labels, all
red reflection. The land cover type was interpreted dynamically for spectral profiles extracted from images acquired at different times for
multi-temporal Landsat images, and interpretations for different time the training sites can be used to enrich the spectral diversity of the train-
were all recorded in the sample set. The interpretation was also sup- ing dataset. For the same purpose, training samples can be spatially ex-
ported by MODIS EVI time series, high resolution imagery, historical im- panded surrounding the center of sample locations, according to their
agery, street views and photos on Google Earth to lower uncertainty. homogeneity.
Beside the acquisition time of the interpreted Landsat image (DOY and In the past, for land cover mapping at larger scale, images were usu-
year) and the spatial size which is homogeneous around the sample ally classified using a model trained by training samples of similar time
(1, 3, 8 or 17 pixels of 30 m), the disturbance of thin clouds, cloud and space, namely Spatial Temporal Sample Search (STS) approach
shadows, relief shadows and snow background are also marked in (Gong et al., 2013). Unexpectedly, a common classification model for
each record of interpretation. As a result, we collected 4107 basic loca- global land cover mapping was proved to have better accuracy (Yu et
tions for training. For each location we interpreted multi-temporal al., 2014b). Therefore, in this study we want to build a spatial and tem-
Landsat images, and got 40,162 temporal interpretation records (with poral common model which is suitable for the classification of any
an average of 10 different temporal interpretation records for each loca- Landsat scene. In order to test the effectiveness of the combination of
tion). Each interpretation record can be spatially expanded depending multi-temporal training samples, we trained the model with only sam-
on the spatial size which is homogeneous around the sample location ples of each month and each season respectively, and used each model
(1, 3, 8 or 17 pixels of 30 m). Finally, we got 1,227,467 pixels of 30 m to classify all validation samples. The accuracies were compared with
in the training dataset. The validation sample set was collected two the accuracy we got when the model was trained by the combination
ways: manual interpretation and field visitation. Unlike training sam- of multi-temporal training samples. The feature set was simplified to
ples which were manually selected in the stratified sampling blocks, only spectral bands of single image in this experiment. According to
validation sample sites were randomly generated by stratified sampling the results (shown in Fig. 4), model trained by the combination of
to ensure an objective accuracy assessment. To get a better sense of local multi-temporal sample set performed better than the ones trained by
land cover and collect more accurate ground-based observations, the sample set of single season or single month. Generally the model

Fig. 7. Classification accuracies of the image acquired in each month by the model trained by training samples of each month, each season, or the whole year (new feature combination
includes multi-temporal NDVI series).
 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185 177

trained by the sample set of each month worked best on the classifica- Features selected for classification include not only spectral data
tion of images acquired around the same time, while its performance and NDVI (Normalized Difference Vegetation Index) as well as
was weak if the time gap is large. This experiment results clearly show NDWI (Normalized Difference Water Index) calculated from the
the advantage of the design of using multi-temporal training sample Landsat scene for classification, but also a multi-season NDVI series
set in this study. derived from all Landsat images acquired on different dates for the

Fig. 8. The logical rules for the correction of seasonal land cover map according to the integrated map.
178 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185

same location to help differentiate land cover classes with seasonal

variation. Synthetic data was generated by NDVI MVC (maximum
value composite) algorithm for each season to minimize cloud con-
tamination, shadow effects, aerosol and water vapor effects
(Holben, 1986). In the calculation of the multi-season NDVI MVC se-
ries for training samples, if there are several interpretation records
(different DOYs) in the same season, records without any distur-
bance would be used with high priority. Two groups of examples
were given to show the differences of NDVI time series derived
from multi-temporal Landsat images among different land cover
types. The samples in each group are very close to each other, shar-
ing similar climate conditions, but their NDVI time series are differ-
ent due to the land cover differences. In Fig. 5, the NDVI of shrubland
begins to increase in autumn and reaches to the peak in winter. The
example of grasslands is on the mountain ridge, and the NDVI main-
tained at about 2000 throughout the year. The example of barren
lands is on the flood plain of a river, with NDVI below 1000 contin-
uously. The example of croplands starts growing when spring
comes, peaks in summer and is harvested in autumn. The spectral
data maybe similar for two land cover types in one season, but
they have different signatures of NDVI time series. Similarly, Fig. 6
shows the difference clearly among three forest types in this area.
Conifer plantations have high and stable NDVI for all seasons,
while broadleaf plantations have lower NDVI in dry summer. Natu-
ral broadleaf forests turn yellow in winter, so their NDVI drop to
the lowest. Even though the calculation of MVC results in loss of in-
formation, it helps to eliminate the disturbance of clouds, and the
temporal pattern is kept basically. We tested many combinations
of features, our final selected one achieved the best accuracies
(shown in Fig. 7), mainly owing to the multi-temporal NDVI series.
The average accuracy of all months without the multi-temporal
NDVI series in the feature set was 73.8%, but this number rose to
83.5% using final selected feature set.
All Landsat images were classified, meaning that for one path/row
location, several land cover maps were produced from images acquired
at different times. The land cover maps were integrated mainly by pixel-
based majority voting, followed by the adjustment according to the
principles of “the greenest” and “the wettest”, i.e., vegetation classes
have the highest priority on deciding the integrated land cover class
(vegetation N water N impervious surfaces N barren lands N snow and
ice N clouds). “Clouds” has the lowest priority. It will be filter out if the
corresponding pixel is categorized into any land cover type which has
higher priority than “Clouds”, and be replaced by the majority type of
the highest priority as its integrated land cover type. If all images
show that this pixel is covered by clouds, it will be kept in the final re-
sults, shown as “unidentified” in the integrated map, accounting for a
very small percentage of the map.
We consider that the integrated land cover map has lower uncer-
tainty than the individual map produced by a single image acquired at
a particular time. To get the seasonal dynamic land cover maps, majority
voting and stacking of the individual maps for each season was
corrected by the logical inference from the integrated land cover map.
In Fig. 8, the logical rules to decide whether the seasonal land cover
class should be replaced by the integrated class or kept in its original
seasonal class or assigned a modified seasonal class are presented. The
vertical axis represents the land cover class on the integrated map, Fig. 9. The integrated land cover map of Chile displayed at the level 1 (plus some level 2
while the horizontal axis represents the land cover class on the map of classes) generalization.
a single season for the same location. Uncertainty maps are also gener-
ated for all seasonal maps using the coding system shown by the circled
numbers in Fig. 8, following the rule that the smaller the number, the 3. Results and discussions
lower the uncertainty of the seasonal map. For example, the integrated
map shows 1 pixel is shrublands, but the majority voting result for this 3.1. Products and validation
pixel in summer is barren, then it should be withered shrublands. The
land cover change affected by life cycle of vegetation and seasonal fluc- An integrated land cover map for circa 2014 at 30 m spatial resolu-
tuations of water supply was depicted by these seasonal dynamic land tion was generated for all three levels in the classification scheme. To
cover maps. display it better, the map is shown primarily at level 1, with some
 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185 179

Table 3
Confusion matrix for the integrated Chilean land cover map presented at level 1 (the bold data in the table represents the number of points for which the predicted label is equal to the
reference label).

Predicted Reference

Crop Forests Grass Shrub Water Impervious Barren Snow Clouds UA (%)

Croplands 50 4 9 1 0 1 0 0 0 77
Forests 2 324 5 20 2 0 0 0 0 92
Grasslands 5 11 184 33 1 0 0 0 0 79
Shrublands 8 50 59 257 4 1 92 0 0 54
Water 0 1 1 0 50 0 5 0 0 88
Impervious 0 0 0 0 0 8 3 0 0 73
Barren 0 0 6 10 0 4 441 1 0 95
Snow/ice 0 0 2 0 0 0 3 29 0 85
Clouds 0 0 0 0 0 0 0 1 0 0
PA (%) 77 83 69 80 88 57 81 94 – 80

level 2 classes like natural forests, plantations, fallow and withered veg- because the urban areas are usually mixed pixels of paved surfaces and
etation in Fig. 9. Accuracy assessment for this map was performed and vegetation. In addition, the impervious surfaces sometimes look spec-
the overall accuracy is 80% based on an independent validation dataset trally similar to barren lands.
containing 1688 sample sites well-distributed in the whole country. The overall accuracy for level 2 classes is 73%. Among the subclasses,
From the confusion matrix of the map presented at level 1 (Table 3), herbaceous croplands (83% for users accuracy and 74% for producers ac-
we could find that all classes were classified with acceptable accuracies. curacy), native forests (85% for users accuracy and 76% for producers ac-
Forests and barren lands have the top users accuracies among all classes, curacy), conifer plantations (88% for users accuracy and 77% for
which are 92% and 95%, respectively. Croplands, grasslands, water, snow producers accuracy), marshlands (76% for users accuracy and 76% for
and ice are classified with users accuracies around or higher than 80%. producers accuracy), sandy and rock areas (94% for users accuracy and
Snow and ice have the best producers accuracy which is 94%, followed 80% for producers accuracy) had good classification performance. The
by forests and barren bands. Shrublands have relatively high producers level 2 maps for these classes are useful for many applications such as
accuracy, but the users accuracy is quite low. Some classified shrublands biodiversity conservation, plantation management, and agriculture
are actually barren lands. This confusion is caused by the fact that many planning and yield estimation. The accuracies for orchards, broadleaf
shrublands in the training sample are comprised of cactus and dry plantations, pastures, natural grasslands are ranging from 60% to 70%.
bushes whose spectral profiles are similar to barren lands. The impervi- However, subclasses of shrublands are poorly classified, with the users
ous surface is the worst classified class, with quite high omission error, accuracies varying from 38% to 49% and the producers from 14% to

Fig. 10. The seasonal land cover map series of Chile displayed at the level 1 (plus some level 2 classes) generalization.
180 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185

Table 4
Overall accuracies for seasonal land cover maps at different levels.

Product level Summer (%) Autumn (%) Winter (%) Spring (%) Integrated (%)

Level 1 corrected 70 70 69 69 80
Level 1 original 49 49 68 47
Level 2 corrected 65 64 64 64 73
Level 2 original 43 44 63 43
Level 3 corrected 53 53 54 53 59
Level 3 original 34 34 53 36

79%. Shrubs and sparse trees mosaic were the worst, highly confused winter and its withering up from winter to summer. The seasonal land
with secondary native forests. cover map series of the north end of the Patagonian Icefield were en-
Seasonal land cover dynamics were reflected in the seasonal maps larged and shown on the lower right corner of each map in Fig. 10, pre-
(Fig. 10). Phenological change of grasses and shrubs, alternation of senting the seasonal expanding and shrinking of snow cover. Since for
farming and fallow, variation of water surface, accumulating and melt- each validation sample site, dynamic land cover type for each DOY
ing of snow cover can all be seen on the seasonal maps. The enlarged was recorded, the seasonal land cover maps can be validated using the
map on the upper left corner of Fig. 10 shows the area in Valparaíso Re- validation dataset. The overall accuracies for the seasonal land cover
gion and Santiago Metropolitan Region with a Mediterranean climate. It maps at different level are shown in Table 4, for maps before and after
is obvious to see the greening up of natural vegetation from summer to correction according to the integrated map. The overall accuracies for
the integrated map are higher than the land cover maps for each season,
and much higher than the seasonal maps without the correction. Due to
the fact that a large proportion of the country has a Mediterranean cli-
mate with wet season in the winter, the accuracies of the original sea-
sonal map from winter images were higher than other seasons
without the correction. This proves that images acquired in the growing
season of vegetation are superior for land cover mapping.
Our result shows the distribution of areas with seasonal land cover
change (see Fig. 11), which accounts for 26.37% of the total area. The
major type of seasonal land cover change varies spatially. In the north,
seasonal change is mainly the sprouting of vegetation in one season
from the barren deserts of other seasons. Another important type of sea-
sonal change in the north is the alternation of salt lakes in summer and
salt flats or dry sands for other seasons. In central Chile, seasonal change
is mainly shrublands and withered shrublands determined by the win-
ter rainy season of a Mediterranean climate. The southern portion of
central Chile has a dominant type of seasonal change from croplands
to bare croplands. On the Andes, the main seasonal change comes
from the snow cover in winter, as well as the wet south.

3.2. Highlights of thematic contents

The high spatial resolution and the detailed classification scheme of

our Chilean land cover products enable users to identify many thematic
contents for their own applications. Some representative examples are
given below.
One of the most typical and important landscapes in Chile is or-
chards, which have been well-identified in our maps. Since the 1990s,
people have been seeking for potential viticulture area outside the Chil-
ean Central Valley. Due to the high evapotranspiration caused by the
abundant sunshine and limited precipitation, they developed their agri-
culture along the rivers that run down from the Andes Mountains. Fig.
12 shows an example of orchards along one of the most important riv-
ers, the Elqui River in the Region of Coquimbo. From the land cover map
in Fig. 12, the impervious surfaces on the upper left corner is La Serena
on ocean terraces. Orchards in this area extend from the Pacific coast
eastward to the Andes Mountains along the Elqui valley, where people
have access to water for irrigation to produce high-quality wine grapes,
table grapes and other fruits. Since the area is at the southern end of the
Atacama Desert, we can see shrublands and dry rocky terrain surround-
ing the river valley. The zoomed-in maps of two areas in the dashed
boxes are shown with higher resolution DigitalGlobe images. On the
Fig. 11. Distribution of area with seasonal land cover change and the major type of left, it is clear to see that the orchards were separated from the herba-
seasonal change. ceous croplands. The orchards climb up to higher altitudes along the
 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185 181

river valley. From the location presented on the right, the orchard valley discrimination of native forests from broadleaf plantations. The planta-
narrows down to the east showing the advantage of the 30 m spatial tions dominate on gentle slopes, while native forests only exist in the
resolution for mapping these features. Field photos taken near this valley between the slopes. The zoomed-in map on the right is a mix of
area are presented in Fig. 12. conifer plantation (left), broadleaf plantation (lower middle) and sec-
Wetlands on the high altitude of the Andean mountains, supplied ondary native forests (right), and they are all mapped correctly. Field
mainly by the outcrops of groundwater, melting glaciers, and lowland photos presented here are pine plantations (left) and eucalyptus planta-
accumulated precipitation, are distinctive ecosystems in the arid high- tions (right), mixed with some shrubs and short native forests.
lands, locally known as “Bofedales”, which is still poorly understood.
These wetlands provide special natural habitats for wildlife and are vul-
nerable to climate change. Our land cover maps provide accurate and 3.3. Spatial variation of accuracy
dynamic spatial information about the wetlands, the lakes and the salt
flats. Fig. 13 shows the map and field photos for part of the Nevado Information on the spatial variation in accuracy is important because
Tres Cruces National Park located in the Atacama Region. On the huge it can indicate regions where more efforts should be spent on land cover
salt flat, referred to as “Salar de Maricunga”, shown in the center of map improvements. The overall accuracies were calculated for the
the map, we can see the lagoons and wetlands. With the help of season- WWF ecoregions. Fig. 15 shows the spatial variation of accuracies for
al land cover maps on the upper right corner in Fig. 13, seasonal varia- level 1 classes, and similar patterns are found for level 2 and level 3 clas-
tion of the lagoon water surface and the phenological change of the ses. The results indicate the overall accuracy decreases gradually from
vegetation can be identified. The lagoons and wetlands are extremely central Chile to both the north and the south.
important for many wildlife species. For example, the Andean flamingos The highest local accuracy is in the ecoregion of Chilean Matorral
shown on the field photos taken in this area, reside in such salt lakes and (83.4% at level 1), characterized by a Mediterranean climate. From the
wetlands during the summer, and migrate to the lower wetlands in the very north part of this ecoregion to the immediate south of La Serana
winter. They are one of the rarest flamingos in the world and vulnerable is a semi-desert area called “Norte Chico” covered mainly by barren
to habitat change. lands, withered shrublands, shrublands and some croplands especially
Native forests and plantations were also differentiated reasonably orchards. The Mediterranean center starts from the immediate south
well. This separation is important as it can provide information for na- of La Serana to the south of the central valley, dominated by shrublands
tive forest conservation and plantation management. Fig. 14 shows an and some sclerophyll forests on moist seaward slopes as natural vegeta-
example near the border of the Bío-Bío Region and the Araucanía Re- tion, and croplands as well as grassland in the central valley as a result of
gion. This region covers the coastal range dominated by plantations on human activities. The class of shrublands occupies the biggest area in
the west and the central valley dominated by croplands and pastures this ecoregion, which is well-classified with a producers accuracy of
on the east. The zoomed-in map on the left demonstrates the 87% and a user's accuracy of 75%.

Fig. 12. Mapping result and field photos of the orchards along the Elqui Valley.
182 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185

Fig. 13. Dynamic land cover maps and field photos of the wetlands and salt flats near Salar de Maricunga.

The land cover map is also more accurate (83.3% for level 1) for the into gravels which is also a major confusion. The wetlands of this area
ecoregion of southern Andean steppe, with an altitudinal zonation of are well-identified whose producers and users accuracies are both
grasslands, shrublands, withered vegetation and barren lands. The bar- higher than 90%.
ren lands and grasslands were classified with better performance (85.3% The Atacama Desert has a lower accuracy (80.3% at level 1) because
and 100% for producers accuracy, 98.3% and 87.5% for users accuracy, re- the validation dataset did not sample enough in the vast desert
spectively), but shrublands are poorly classified. The large proportion of completely devoid of life. There are only 51 sample units in this region,
barren lands in this ecoregion raised the overall accuracy. and there is a confusion between barren lands and withered shrublands.
The overall accuracy for the ecoregion of Valdivian temperate forests The Magellanic subpolar forests and Patagonian steppe have the
at level 1 is 82.9%. The dominant land cover type is forests, which are lowest accuracies of 64.8% and 66.7%, respectively. This is due to lack
mainly plantations along the coastal range and native forests on the of cloud free images over these regions. In these regions, forests are
Andes Mountains. The native forest and plantations are categorized short resulting in more confusion with shrublands.
with producers accuracies of 81.7% and 79.4%, users accuracies of
88.4% and 80.6%. The major land cover type for the central valley is pas- 4. Conclusions and perspectives
tures, whose producers accuracy is 73.9% and users accuracy is 59.6%,
mainly confused with natural grasslands. A 30 m spatial resolution Chilean land cover map and a seasonal land
The Central Andean dry puna is a unique ecoregion located in the cover map series were produced for circa 2014 primarily using Landsat
Andean High plateau, which includes mainly barren lands, dry 8 imagery. The classification scheme is more detailed than existing land
shrublands and grasslands, and interspersed with some wetlands cover databases. The classes derived in this study are able to support
along the streams. The overall accuracy is 81.4% at level 1, but the dry more applications in biodiversity conservation, land surface modelling,
shrublands are confused with barren lands in this ecoregion, due to resource management and planning. This is the first nationwide land
their similar spectral characteristics. Some salt flats were categorized cover map for Chile through close international collaboration between
 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185 183

Fig. 14. Mapping result and field photos of the native forests and plantations in Bío-Bío Region.

local experts and researchers in China and the US. Field works by the in- training set outside the mask, while putting more training samples of
ternational team helped improving the scheme design, training sample impervious surfaces inside the mask, and merge these two layer for a
interpretation and validation. The overall accuracy reached 80% for level final result. In addition, microwave data might be useful over the subpo-
1 product and 73% for level 2 thanks to the integration of multi-seasonal lar areas that are frequently covered by clouds. For example, Advanced
data. This study is among the first ones that generate seasonal land Land Observing Satellite (ALOS) Phased Array type L-band Synthetic
cover dynamics. Our previous dynamics land cover mapping efforts Aperture Radar (PALSAR) data has been proved effective in some re-
with medium resolution satellite data are restricted to a lake or munic- searches (Zhu et al., 2012).
ipality area (Dronova et al., 2015; Wang et al., 2015). Here we extended The multi-temporal NDVI series we used now in the feature set is the
this capability to a whole nation. The seasonal land cover maps for Chile MVC result, in order to eliminate the disturbance of cloud contamina-
are also validated. Through correction according to the integrated map, tion and keep the feature vector in the same length. Sometimes the
it is possible to substantially improve accuracies of seasonal land cover MVC result may hide some patterns in the seasonal variation. We can
maps. The analysis of spatial variation of accuracy among the ecoregions take not only MVC but also other metrics for each season, for example
indicates that the accuracy for land cover mapping decreases gradually the upper and lower quartiles. Experiments should be designed for fea-
from central Chile to both north and south, providing useful information ture selection. The integration strategy of multi-seasonal mapping as
for directing future efforts in Chile. well as the logic rules for the seasonal map correction could also be fur-
Multi-temporal images play a vital role in improving our mapping ther refined. The post-classification strategy in our study is simple and
process, and they work in three ways together. Experiments showed fast, and it can be improved by estimating the probability of all alterna-
that mapping accuracies were improved by (1) combining multi-tem- tives and analyzing the uncertainty using more statistical methods
poral training samples collected from multi-temporal images, (2) (McRoberts, 2010; Olofsson et al., 2012; Sexton et al., 2015).
adding a four season NDVI series to the feature set, and (3) integrating The mapping result reported here is only for 1 year around 2014.
the classification results of many multi-temporal images. Long-term land cover dynamics are important for monitoring land
The accuracy for land cover mapping in Chile could be further im- cover changes which can be used in many applications, such as studies
proved by the inclusion of additional ancillary data. DMSP Operational on fire disturbance (Carmona, González, Nahuelhual, & Silva, 2012), re-
Linescan System (OLS) nighttime stable light (NTL) datasets could be placement of native forest by exotic plantations (Little, Lara, McPhee, &
used to differentiate the urban area, which is poorly classified in the cur- Urrutia, 2009), deforestation and habitat change (Echeverria et al.,
rent map. Since the urban area is always a mosaic of vegetation and im- 2006; Vergara, Perez-Hernandez, Hahn, & Soto, 2013). The training
pervious surface, training samples of impervious surfaces are usually dataset includes samples spatially distributed in the whole country,
not pure pixels, leading to serious confusion between impervious sur- temporally distributed throughout a year, and covering different land
faces and other land cover types. Though the night light data has cover types. We believe that we have a relatively exhaustive training
lower spatial resolution, it can be used as a mask for potential urban dataset that could be used to train a universal model for the classifica-
area, i.e. removing training samples of impervious surfaces from the tion of any image newly acquired for Chile. To produce a long-term
184 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185

Arvor, D., Jonathan, M., Meirelles, M. S. P., Dubreuil, V., & Durieux, L. (2011). Classification
of modis evi time series for crop mapping in the state of Mato Grosso, Brazil.
International Journal of Remote Sensing, 32, 7847–7871.
Blanco, P. D., Colditz, R. R., Saldaña, G. L., Hardtke, L. A., Llamas, R. M., Mari, N. A., Fischer, A.
, Caride, C., Aceñolaza, P. G., Del Valle, H. F., et al. (2013). A land cover map of Latin
America and the Caribbean in the framework of the serena project. Remote Sensing
of Environment, 132, 13–31.
Brisco, B., & Brown, R. (1995). Multidate sar/tm synergism for crop classification in west-
ern Canada. Photogrammetric Engineering and Remote Sensing, 61, 1009–1014.
Carmona, A., González, M. E., Nahuelhual, L., & Silva, J. (2012). Spatio-temporal effects of
human drivers on fire danger in Mediterranean Chile. Bosque, 33, 321–328.
Cartwright, D., & Gaston, C. (2002). Chile's forest products markets-a plantation success
story. Forest products annual market review.
Dai, Y., Zeng, X., Dickinson, R. E., Baker, I., Bonan, G. B., Bosilovich, M. G., Denning, A. S.,
Dirmeyer, P. A., Houser, P. R., Niu, G., et al. (2003). The common land model.
Bulletin of the American Meteorological Society, 84, 1013–1023.
Di Gregorio, A. (2005). Land cover classification system: Classification concepts and user
manual: LCCS. Vol. 8, Food & Agriculture Org.
Dong, Z., Wang, Z., Liu, D., Song, K., Li, L., Jia, M., & Ding, Z. (2014). Mapping wetland areas
using landsat-derived ndvi and lswi: A case study of west Songnen plain, Northeast
China. Journal of the Indian Society of Remote Sensing, 42, 569–576.
Dronova, I., Gong, P., Wang, L., & Zhong, L. (2015). Mapping dynamic cover types in a large
seasonally flooded wetland using extended principal component analysis and object-
based classification. Remote Sensing of Environment, 158, 193–206.
Echeverria, C., Coomes, D., Salas, J., Rey-Benayas, J. M., Lara, A., & Newton, A. (2006). Rapid
deforestation and fragmentation of Chilean temperate forests. Biological Conservation,
130, 481–494.
Espinosa, M., Acuña, E., Cancino, J., Muñoz, F., & Perry, D. A. (2005). Carbon sink potential
of radiata pine plantations in Chile. Forestry, 78, 11–19.
Eva, H. D., Belward, A. S., De Miranda, E. E., Di Bella, C. M., Gond, V., Huber, O., ... Fritz, S.
(2004). A land cover map of South America. Global Change Biology, 10, 731–744.
Fagan, M. E., DeFries, R. S., Sesnie, S. E., Arroyo-Mora, J. P., Soto, C., Singh, A., ... Chazdon, R.
L. (2015). Mapping species composition of forests and tree plantations in northeast-
ern Costa Rica with an integration of hyperspectral and multitemporal landsat imag-
ery. Remote Sensing, 7, 5660–5696.
Friedl, M. A., Sulla-Menashe, D., Tan, B., Schneider, A., Ramankutty, N., Sibley, A., &
Huang, X. (2010). Modis collection 5 global land cover: Algorithm refinements
and characterization of new datasets. Remote Sensing of Environment, 114,
168–182.
Giri, C., & Long, J. (2014). Land cover characterization and mapping of South America for
the year 2010 using landsat 30 m satellite data. Remote Sensing, 6, 9494–9510.
Gong, P. (2012). Remote sensing of environmental change over China: A review. Chinese
Science Bulletin, 57, 2793–2801.
Gong, P., Wang, J., Yu, L., Zhao, Y., Zhao, Y., Liang, L., Niu, Z., Huang, X., Fu, H., Liu, S., et al.
(2013). Finer resolution observation and monitoring of global land cover: First map-
ping results with landsat tm and etm + data. International Journal of Remote Sensing,
34, 2607–2654.
Guerschman, J., Paruelo, J., Bella, C. D., Giallorenzi, M., & Pacin, F. (2003). Land cover clas-
sification in the argentine pampas using multi-temporal landsat tm data.
International Journal of Remote Sensing, 24, 3381–3402.
Hansen, M., DeFries, R., Townshend, J. R., & Sohlberg, R. (2000). Global land cover classi-
fication at 1 km spatial resolution using a classification tree approach. International
Journal of Remote Sensing, 21, 1331–1364.
Fig. 15. Spatial variation of overall accuracy among different ecoregions. Henebry, G. M. (1993). Detecting change in grasslands using measures of spatial depen-
dence with landsat tm data. Remote Sensing of Environment, 46, 223–234.
Hess, L. L., Melack, J. M., Novo, E. M., Barbosa, C. C., & Gastil, M. (2003). Dual-season map-
ping of wetland inundation and vegetation for the central amazon basin. Remote
Sensing of Environment, 87, 404–428.
Hojas Gascon, L., Eva, H., Gobron, N., Simonetti, D., & Fritz, S. (2012). The application of
dynamic product for the past, images of Landsat 5 Surface Reflectance medium-resolution meris satellite data for continental land-cover mapping over
CDR datasets were also prepared back to the 1980s. We will test the South America—Results and caveats. Remote sensing of land use and land cover: Prin-
training sample from Landsat 8 images to classify Landsat 5 images fol- ciples and applications. Boca Raton, Fl: CRC/Taylor & Francis.
Holben, B. N. (1986). Characteristics of maximum-value composite images from temporal
lowing some consistency checks. We expect to produce a historical avhrr data. International Journal of Remote Sensing, 7, 1417–1434.
long-term land cover dynamics multi-seasonal dataset for Chile in the Homer, C., Dewitz, J., Fry, J., Coan, M., Hossain, N., Larson, C., ... Wickham, J. (2007). Com-
near future. The land cover data presented in this paper will be open ac- pletion of the 2001 national land cover database for the counterminous United States.
Photogrammetric Engineering and Remote Sensing, 73, 337.
cess on data.ess.tsinghua.edu.cn. Jung, M., Henkel, K., Herold, M., & Churkina, G. (2006). Exploiting synergies of global land
cover products for carbon cycle modeling. Remote Sensing of Environment, 101, 534–553.
LaFontaine, J. H., Hay, L. E., Viger, R. J., Regan, R. S., & Markstrom, S. L. (2015). Effects of
Acknowledgements climate and land cover on hydrology in the Southeastern US: Potential im-
pacts on watershed planning. JAWRA Journal of the American Water Resources
Association.
This research is funded by a project (Development of advanced re-
Lanfranco, D., & Dungey, H. (2001). Insect damage in eucalyptus: A review of plantations
mote sensing methods for mapping and managing plant species diversi- in Chile. Austral Ecology, 26, 477–481.
ty in Mediterranean Forests of Chile and California, 2013–2014) in UC Liang, L., Xu, B., Chen, Y., Liu, Y., Cao, W., Fang, L., ... Li, W. (2010). Combining spatial-tem-
Berkeley-Chile Seed Funds (grant number: 77477) and a Meteorological poral and phylogenetic analysis approaches for improved understanding on global
h5n1 transmission. PloS One, 5, e13575.
Public Benefit project of China (grant number: GYHY201506010). Little, C., Lara, A., McPhee, J., & Urrutia, R. (2009). Revealing the impact of forest exotic
plantations on water yield in large scale watersheds in south-central Chile. Journal
of Hydrology, 374, 162–170.
References Liu, J., Chen, J., Cihlar, J., & Park, W. (1997). A process-based boreal ecosystem productivity
simulator using remote sensing inputs. Remote Sensing of Environment, 62, 158–175.
Altamirano, A., Aplin, P., Miranda, A., Cayuela, L., Algar, A. C., & Field, R. (2013). High rates Lo, T. H., Scarpace, F. L., & Lillesand, T. M. (1986). Use of multitemporal spectral profiles in
of forest loss and turnover obscured by classical landscape measures. Applied agricultural land-cover classification. Photogrammetric Engineering and Remote
Geography, 40, 199–211. Sensing (USA).
 Y. Zhao et al. / Remote Sensing of Environment 183 (2016) 170–185 185

McRoberts, R. E. (2010). Probability- and model-based approaches to inference for pro- Sun, F., Zhao, Y., Gong, P., Ma, R., & Dai, Y. (2014). Monitoring dynamic changes of global
portion forest using satellite imagery as ancillary data. Remote Sensing of land cover types: Fluctuations of major lakes in China every 8 days during 2000–
Environment, 114, 1017–1025. 2010. Chinese Science Bulletin, 59, 171–189.
Moreira-Muñoz, A. (2011). Plant geography of Chile. vol. 5, Springer Science & Business Toro, J., & Gessel, S. (1999). Radiata pine plantations in Chile. In J. Boyle, J. Winjum, K.
Media. Kavanagh, & E. Jensen (Eds.), Planted forests: Contributions to the quest for sustainable
Munro, D., & Touron, H. (1997). The estimation of marshland degradation in southern societies. Of Forestry Sciences, Vol. 56. (pp. 393–404). Springer Netherlands (URL:
Iraq using multitemporal landsat tm images. International Journal of Remote Sensing, http://dx.doi.org/10.1007/978-94-017-2689-4_25 . doi: 10.1007/978–94–017-2689-
18, 1597–1606. 4_25).
Nahuelhual, L., Carmona, A., Lara, A., Echeverra, C., & González, M. E. (2012). Land-cover Townshend, J. R., Justice, C., & Kalb, V. (1987). Characterization and classification of South
change to forest plantations: Proximate causes and implications for the landscape American land cover types using satellite data. International Journal of Remote Sensing,
in south-central Chile. Landscape and Urban Planning, 107, 12–20. 8, 1189–1207.
Olofsson, P., Stehman, S. V., Woodcock, C. E., Sulla-Menashe, D., Sibley, A. M., Newell, J. D., Tucker, C. J., Townshend, J. R., & Goff, T. E. (1985). African land-cover classification using
... Herold, M. (2012). A global land-cover validation data set, part i: Fundamental de- satellite data. Science, 227, 369–375.
sign principles. International Journal of Remote Sensing, 33, 5768–5788. Vergara, P. M., Perez-Hernandez, C. G., Hahn, I. J., & Soto, G. E. (2013). Deforestation in
Özesmi, U., & Mitsch, W. J. (1997). A spatial habitat model for the marsh-breeding red- central Chile causes a rapid decline in landscape connectivity for a forest specialist
winged blackbird (Agelaius phoeniceus l.) in coastal Lake Erie wetlands. Ecological bird species. Ecological Research, 28, 481–492.
Modelling, 101, 139–152. Wang, J., Zhao, Y., Li, C., Yu, L., Liu, D., & Gong, P. (2015). Mapping global land cover in
Panayotou, T., & Ashton, P. (1992). Not by timber alone: Economics and ecology for sustain- 2001 and 2010 with spatial-temporal consistency at 250 m resolution. ISPRS Journal
ing tropical forests. Island Press. of Photogrammetry and Remote Sensing, 103, 38–47.
Pauleit, S., & Duhme, F. (2000). Assessing the environmental performance of land cover Wolter, P. T., Mladenoff, D. J., Host, G. E., & Crow, T. R. (1995). Improved forest classifica-
types for urban planning. Landscape and Urban Planning, 52, 1–20. tion in the northern lake states using multi-temporal landsat imagery.
Pearson, R. G., Dawson, T. P., & Liu, C. (2004). Modelling species distributions in Britain: A Photogrammetric Engineering and Remote Sensing, 61, 1129–1144.
hierarchical integration of climate and land-cover data. Ecography, 27, 285–298. Yang, J., Gong, P., Fu, R., Zhang, M., Chen, J., Liang, S., ... Dickinson, R. (2013). The role of
Peña, M., & Brenning, A. (2015). Assessing fruit-tree crop classification from landsat-8 satellite remote sensing in climate change studies. Nature Climate Change, 3, 875–883.
time series for the Maipo valley, Chile. Remote Sensing of Environment, 171, 234–244. Yu, L., Liang, L., Wang, J., Zhao, Y., Cheng, Q., Hu, L., ... Gong, P. (2014a). Meta-discoveries
Potapov, P., Turubanova, S., Tyukavina, A., Krylov, A., McCarty, J., Radeloff, V., & Hansen, M. from a synthesis of satellite-based land-cover mapping research. International Journal
(2015). Eastern Europe's forest cover dynamics from 1985 to 2012 quantified from of Remote Sensing, 35, 4573–4588. http://dx.doi.org/10.1080/01431161.2014.930206
the full landsat archive. Remote Sensing of Environment, 159, 28–43. (URL: http://www.tandfonline.com/doi/abs/10.1080/01431161.2014.930206).
Schriever, J. R., & Congalton, R. G. (1995). Evaluating seasonal variability as an aid to Yu, L., Shi, Y., & Gong, P. (2015). Land cover mapping and data availability in critical ter-
cover-type mapping from landsat thematic mapper data in the northeast. restrial ecoregions: A global perspective with landsat thematic mapper and enhanced
Photogrammetric Engineering and Remote Sensing, 61, 321–327. thematic mapper plus data. Biological Conservation, 190, 34–42.
Schulz, J. J., Cayuela, L., Echeverria, C., Salas, J., & Benayas, J. M. R. (2010). Monitoring land Yu, L., Wang, J., Li, X., Li, C., Zhao, Y., & Gong, P. (2014b). A multi-resolution global land
cover change of the dryland forest landscape of central Chile (1975–2008). Applied cover dataset through multisource data aggregation. Science China Earth Sciences,
Geography, 30, 436–447. 57, 2317–2329.
Senf, C., Leitão, P. J., Pflugmacher, D., van der Linden, S., & Hostert, P. (2015). Mapping land Zhao, Y., Gong, P., Yu, L., Hu, L., Li, X., Li, C., Zhang, H., Zheng, Y., Wang, J., Zhao, Y., et al.
cover in complex mediterranean landscapes using landsat: Improved classification (2014). Towards a common validation sample set for global land-cover mapping.
accuracies from integrating multi-seasonal and synthetic imagery. Remote Sensing International Journal of Remote Sensing, 35, 4795–4814.
of Environment, 156, 527–536. Zhong, L., Gong, P., & Biging, G. S. (2012). Phenology-based crop classification algorithm
Sexton, J. O., Noojipady, P., Anand, A., Song, X. -P., McMahon, S., Huang, C., ... Townshend, J. and its implications on agricultural water use assessments in California's central val-
R. (2015). A model for the propagation of uncertainty from continuous estimates of ley. Photogrammetric Engineering & Remote Sensing, 78, 799–813.
tree cover to categorical forest cover and change. Remote Sensing of Environment, Zhong, L., Gong, P., & Biging, G. S. (2014). Efficient corn and soybean mapping with tem-
156, 418–425. poral extendability: A multi-year experiment using landsat imagery. Remote Sensing
Squeo, F. A., Estévez, R. A., Stoll, A., Gaymer, C. F., Letelier, L., & Sierralta, L. (2012). Towards of Environment, 140, 1–13.
the creation of an integrated system of protected areas in Chile: Achievements and Zhu, Z., Woodcock, C. E., Rogan, J., & Kellndorfer, J. (2012). Assessment of spectral, polar-
challenges. Plant Ecology & Diversity, 5, 233–243. imetric, temporal, and spatial dimensions for urban and peri-urban land cover classi-
Stone, T. A., Schlesinger, P., Houghton, R. A., & Woodwell, G. M. (1994). Map of the vege- fication using landsat and sar data. Remote Sensing of Environment, 117, 72–82.
tation of South America based on satellite imagery. Photogrammetric Engineering and
Remote Sensing, 60.