Environmental Science and Pollution Research (2019) 26:3621–3635

https://doi.org/10.1007/s11356-018-3824-y

RESEARCH ARTICLE

Mapping terrestrial oil spill impact using machine learning random

forest and Landsat 8 OLI imagery: a case site within the Niger Delta
region of Nigeria
Mohammed S. Ozigis 1,2 & Jorg D. Kaduk 1 & Claire H. Jarvis 1

Received: 18 September 2018 / Accepted: 21 November 2018 / Published online: 7 December 2018
# The Author(s) 2018

Abstract
Terrestrial oil pollution is one of the major causes of ecological damage within the Niger Delta region of Nigeria and has caused a
considerable loss of mangroves and arable croplands since the discovery of crude oil in 1956. The exact extent of landcover loss
due to oil pollution remains uncertain due to the variability in factors such as volume and size of the oil spills, the age of oil, and
its effects on the different vegetation types. Here, the feasibility of identifying oil-impacted land in the Niger Delta region of
Nigeria with a machine learning random forest classifier using Landsat 8 (OLI spectral bands) and Vegetation Health Indices is
explored. Oil spill incident data for the years 2015 and 2016 were obtained from published records of the National Oil Spill
Detection and Response Agency and Shell Petroleum Development Corporation. Various health indices and spectral wavelengths
from visible, near-infrared, and shortwave infrared bands were fused and classified using the machine learning random forest
classifier to distinguish between oil-free and oil spill–impacted landcover. This provided the basis for the identification of the best
variables for discriminating oil polluted from unpolluted land. Results showed that better results for discriminating oil-free and oil
polluted landcovers were obtained when individual landcover types were classified separately as opposed to when the full study
area image including all landcover types was classified at once. Similarly, the results also showed that biomass density plays a
significant role in the characterization and classification of oil contaminated and oil-free pixels as tree cover areas showed higher
classification accuracy compared to cropland and grassland.

Keywords Oil spill . Vegetation health indices . Spectral bands . Random forest . Variable importance . Landcover

Introduction result from underground and surface pipeline leakages, sabo-

tage, and operational failure, as well as transport of oil slicks
An oil spill is the discharge of petroleum hydrocarbon prod- from sea to land (Taheri 2012). Oil can damage vegetation
ucts into marine or terrestrial ecosystem. Terrestrial spills through several mechanisms, such as the ingestion and ab-
sorption of toxic compounds through the biota’s respiratory
Responsible editor: Marcus Schulz structures (Joel and Amajuoyi 2009; Mendelssohn et al.
2012), coating and smothering which affects temperature ad-
* Mohammed S. Ozigis aptation, and gas regulation as well as other life-supporting
[email protected]; [email protected]; processes (Mendelssohn et al. 2012). On shore, oil spill con-
[email protected] tamination has the potential of increasing erosion and loss of
salt marsh due to oil-induced plant mortality (Khanna et al.
Jorg D. Kaduk
[email protected] 2013) and the longer oil resides on land, the greater the impact
and slower the recovery (Gundlach and Hayes 1978; Jackson
Claire H. Jarvis
et al. 1989; Khanna et al. 2013). This results from direct im-
[email protected]
pacts of hydrocarbon crude oil on plant metabolism as well as
1
Department of Geography, University of Leicester, Leicester, United indirect impacts through disruption of plant-water relation-
Kingdom ships and reduced gas exchange between atmosphere and
2
Department of Strategic Space Applications, National Space soil (Hester and Mendelssohn 2000; Khanna et al. 2013;
Research and Development Agency (NASRDA), Abuja, Nigeria Pezeshki et al. 2000).
3622 Environ Sci Pollut Res (2019) 26:3621–3635

In Nigeria, the effects of oil exploration are particularly can provide a basis for assessing future terrestrial based oil
glaring in the Niger Delta. Reduced food productivity, dam- spill impacts and how the inter landcover variability of oil
ages to the subsistence economy, habitat distortion, epidemic polluted and oil-free landcover types contribute to a general
outbreaks, and general social instability are among the numer- landcover change pattern. Furthermore, the effective dis-
ous negative impacts that crude oil exploitation has had in the crimination of oil polluted and oil-free landcovers can pro-
Niger Delta (Onwurah et al. 2007). The Nigerian vide information on the location of oil pipeline leakages
Conservation Foundation in a study in 2006 put the figure and the extent of land area affected by oil in regions with
for oil spilt, onshore and offshore, at 9 to 13 million barrels limited accessibility. This mapping can also provide useful
of oil over the past 50 years. This has massively threatened the landcover discriminatory maps for timely intervention in
well-being of the people (Nriagu 2011). Onwurah et al. (2007) oil spill prone areas, as well as a basis for formulating
noted that a good percentage of oil spills that occurred on the mitigation and remediation strategies before irreversible
dry land between 1978 and 1979 in Nigeria affected farmlands damage is done to the ecosystem. In the long term, howev-
in which crops such as rice, maize, yams, cassava, and plan- er, this approach can also be used to formulate robust and
tain were lost. Similarly, findings from the studies conducted transferable image processing models which can be used to
by the United Nation Environmental Programme (UNEP) in track future terrestrial oil spills leveraging on the pool of
2011 in the Niger Delta suggest that residents are exposed to spectral library generated.
elevated levels of petroleum hydrocarbon in contaminated Some studies have tried to reduce the confusion between
drinking water and outdoor air which posed a serious threat classes by implementing spectral space delineation to obtain
to their health (UNEP 2011). pure image training samples specific to each class to generate
Detecting oil spill through remote sensing is frequently the accurate maps (Aplin and Atkinson 2001; Arif et al. 2015;
basis for establishing the impact of oil pollution near shore, Arroyo et al. 2010; MacLachlan et al. 2017; Tsutsumida
marshes, and mudflat ecosystems. Common techniques used et al. 2016).
for oil spill detection include image spectroscopy (Khanna Generally, two fundamental types of image processing
et al. 2013; Kokaly et al. 2013) and field spectroscopy methodologies exist, parametric and non-parametric algo-
(Mishra et al. 2012), broadband Vegetation Health Indices rithms (Li et al. 2013). While the first is dependent on the
(Adamu et al. 2015; Arellano et al. 2015; Noomen et al. characteristic nature of input variables with respect to statis-
2015), narrowband vegetation indices (Arellano et al. 2015; tical distribution, probability, and clustering of pixel values,
Noomen et al. 2015), and recently airborne SAR polarimetry the non-parametric methods do not require variables to fol-
(Ramsey et al. 2015; Ramsey III et al. 2011; Ramsey et al. low a particular statistical distribution and they also have the
2014). Results from satellite image processing with emphasis ability of discretely handling problems of noise, model
on vegetation health are particularly useful in assessing the fitting, and relatively lower computational demands than
impact of oil on terrestrial mangrove and swamp ecosystems other classification approaches. Several on shore oil spill
as well as fragile near-shore marsh vegetation (Adamu et al. studies have used decision tree algorithms for the
2016; Khanna et al. 2013; Kokaly et al. 2013; Mendelssohn assessment of oil contamination on mangrove and
et al. 2012; Mishra et al. 2012; Noomen et al. 2015; Onwurah marshland. Giri et al. (2011) used a decision tree classifier
et al. 2007; Ramsey et al. 2015; Ramsey III et al. 2011; Shi et al. based on a univariate decision tree (C45.5) algorithm to
2007; Sun et al. 2016; Zabbey and Uyi 2014). This is because classify Landsat and Airborne photography of the
of the toxicity of crude oil and its potential to alter the biophys- Louisiana mangroves. Emphasis was on depicting the
ical and biochemical processes in plants and ecosystem com- spatiotemporal characteristics of ecosystem shifts, in terms
munity. However, most studies in oil spill impact assessment of expansion, retraction, and disappearance. Khanna et al.
have focused on detecting the phenomenon without necessarily (2013) also used a binary decision tree based on vegetation
establishing the extent of the impact of these obnoxious com- index, angle index, and depth of oil absorption to produce a
pounds on the adjoining landcover. Attempts have also been classification map for six classes, oiled soil, oiled dry vege-
made to map landcover changes as a result of the long-term tation, oil-free soil, oil-free dry vegetation, green vegetation,
impact of hydrocarbon on plant communities (Ayanlade and and water to assess oil impact on marshland vegetation of
Howard 2016; Kuenzer et al. 2014; Ochege et al. 2017). A the Louisiana coast. However, little attempts have been
significant number of studies have primarily focused on made to assess the functionality of random forest classifica-
assessing general changes on mangrove fields over time with- tion algorithms for discriminating oil-impacted landcover
out specific efforts to distinguish between the healthy compo- from oil-free landcover at a broader scale. The robust appli-
nents (oil-free) and oil-impacted landcover component, and cation of random forest in the extraction of precise details
how the observed trends affect the broader landcover change. from remotely sensed data has been demonstrated in several
This study focuses explicitly on distinguishing and map- studies (Du et al. 2015; Jhonnerie et al. 2015; Juel et al.
ping oil-free and oil-impacted landcovers separately. This 2015; Liu et al. 2014).
Environ Sci Pollut Res (2019) 26:3621–3635 3623

This study aims to Data

& Explore the potential of the non-parametric random forest Three datasets were used in this research: oil spill incident
machine learning classifier to discriminate pixels of oil data, satellite image (Landsat 8, Operational Land Imager),
polluted landcover from oil-free landcover types within and the landcover data.
the Niger Delta region of Nigeria using Landsat 8 visible,
near-infrared, and shortwave infrared bands and derived
Oil spill incident data
Vegetation Health Indices
& Identify the variables that provide most information for this
The oil spill dataset was obtained from two published sources,
discrimination using this non-parametric method, as several
the Shell Petroleum Development Corporation (SPDC)
studies (Adamu et al. 2015, 2016, 2018; Khanna et al. 2013;
https://www.shell.com.ng/sustainability/environment/oil-
Zhu et al. 2013) have tested the sensitivity of some of these
spills.html and the National Oil Spill Detection and Response
variables to detect oil spill using parametric methods
Agency (NOSDRA) https://oilspillmonitor.ng/. The
& Highlight the possible reduction of confusion between
NOSDRA is a government agency tasked with capturing all
classes by implementing subset classification for the sep-
oil spill incidents both in marine and terrestrial realms across
arate landcover types of cropland, grassland, and tree cov-
the country.
er areas is demonstrated

Landcover data

Materials and methods The landcover map for the African continent produced by the
European Space Agency Climate Change Initiative 2016 was
The study area used in this study (http://2016africalandcover20m.esrin.esa.
int/). The product contains 10 classes for different landcover
The study area defined by four corner coordinates of longitude categories including built-up areas, waterbody, and various
6.957° E latitude 5.025° N, longitude 7.247° E latitude 5.025° vegetation types produced from 20-m-high spatial resolution
N, longitude 6.96° E latitude 4.795° N, and longitude 7.254° Sentinel-2A image over Africa. The tile information covering
E latitude 4.804° N covers 1320 km2 within the Niger Delta the study area was downloaded, subsetted, and used for the
region of Nigeria (Fig. 1). It cuts across Abia and Rivers establishment of appropriate landcover types for the study
States. To the far west corner is the Ukwa West Local area. The major landcover categories used in this study were
Government Area of Abia State and to the easterly corner cropland, grassland, and tree cover areas (TCA). Features
are Ikwerre, Obio/Akpor, Eberi/Omumma, Oyigbo, Eleme, such as built-up areas, waterbody, and baresurface were ex-
and Port Harcourt Local Government Area of Rivers state. cluded from this study as most oil pipelines and the

Fig. 1 Location of the study area

within Nigeria in Africa. The
brown-colored area represents
Nigeria within the African
continent, while the green-colored
area is the oil-producing Niger
Delta region. The image shows
the precise extent of the study area
(Source: ESRI, ArcGIS Base Map
Image, provided by Digital
Globe, GeoEye, and Airbus)
3624 Environ Sci Pollut Res (2019) 26:3621–3635

corresponding spill incidents occur on terrestrial vegetation landcover class (cropland, grassland, and TCA), spill inci-
classes. Thus, their exclusion reduced artifacts and dents located within the landcover classes were identified.
misclassification. This provided the various training and validation sites for
the identification of oil-impacted (polluted landcover) classes.
Landsat 8: OLI image
& Selecting non-polluted sites for the different landcover
The Landsat 8 (OLI data) for the year 2016 was downloaded
from the USGS website (earthexplorer.usgs.gov). The image Non-polluted sample sites are necessary in this study for
acquired was a Landsat surface reflectance higher-level data two main reasons: first, for the identification of oil-free (non-
product processed using the Landsat surface reflectance code polluted) landcover types within the study area and secondly
(LaSRC). The LaSRC makes use of the coastal aerosol band for an effective discrimination between pixels of oil-free and
to perform aerosol inversion tests using auxiliary climate data oil spill–impacted landcovers. Proximity analysis as sug-
from MODIS and a unique radiative transfer model (Roy et al. gested by (Obida et al. 2018; Park et al. 2016; Whanda et al.
2014). Additionally, LaSRC hardcodes the view zenith angle 2016) provided the basis for the selection of the polluted and
to B0^ and solar zenith which are used for calculations as part oil-free vegetation pixels. The minimum rule was set that all
of the atmospheric correction process. The image acquired non-polluted sites must be located at least 600 m away from
and used for this study, acquired on the 6th of December, is all polluted sites based on the maximum area of spill recorded.
a post spill dry season image with little to no cloud cover, This resulted in an 800 m buffer ring around all existing spill
aerosol, and haze effect. Images between the months of points, which avoided any overlap with any likely spill-
March and November had significant cloud cover due to the impacted area. The procedure ensured that sample sites select-
wet season. ed for the respective oil-free landcover are reasonably well-
spaced from the oil polluted sites. Thereafter, the training sites
Methods for the non-polluted landcover categories were selected at ran-
dom outside the buffer ring established. Furthermore, specif-
Sampling regime ically only healthy vegetation as inferred from high-resolution
Google Earth image was chosen.
& Spill incident harmonization
& Pixel selection using buffer analysis
The oil spill data harmonization sought to integrate and
expand the oil spill database for this research. The harmoni- Following the reconciliation and extraction of the oil spill
zation operation was carried out by overlaying both datasets points and the non-polluted sites respectively according to
(NOSDRA and SHELL) in a GIS environment. Points with their respective landcover classes (cropland, grassland, and
repeated information as a result of duplicate capture and mul- TCA), the points were then sub-divided into two categories
tiple spill incidents over the years were identified and marked. for training and validation purpose. Sixty percent of the points
Duplicates (in most cases the SPDC data) were deleted since for individual landcover category were randomly selected for
the dataset provided by NOSDRA is all encompassing as the training, while the other 40% were set aside for validation in
government’s regulatory agency with the responsibility of post classification accuracy assessment. Table 1 shows the
documenting all spill incidents. The spill information relating distribution of the polluted spill sites and oil-free sites accord-
to volume, size, and date of spill was checked, as this provided ing to their respective landcover classification schemes. To
the basis for tracking the spill intensity on the different this end, 30 m buffer ring polygons were established around
landcover types. The minimum area covered by the spill data all the training sites to ensure that only adjacent pixels within
used for this exercise is 1000 sqm, which is greater than a
single Landsat image pixel of 900 sqm. This is to ensure that Table 1 Total number of sites used for calibrating and validating the
pixels used for training, testing, and validation of the final random forest classification
model as well as the image classification have dominant spec-
Class label Number of spill sites
tral reflectance of a typical oil polluted site.
Non-polluted cropland 41
& Assignment of spill incidents to landcover Non-polluted grassland 27
Non-polluted tree cover areas 25
The assignment of oil spill incidents to the corresponding Polluted cropland 44
landcover categories is an important step in this study, as the Polluted grassland 26
RF algorithm would rely on the spectral signatures provided Polluted tree cover areas 26
by these training sites to build a robust model. For each
Environ Sci Pollut Res (2019) 26:3621–3635 3625

the high consequence area close to the point of impact are seeks to optimize training samples by randomly selecting sam-
selected specially for the polluted sites (Alexakis et al. 2016; ples to split each node in the decision trees to maximize pre-
Whanda et al. 2016). diction accuracy. This offers the opportunity of including
many variables in a single classification operation, which in
Image preprocessing turn should contribute positively to the prediction of the final
class. A list of variable importance and their contribution to-
As the Landsat surface reflectance higher-level data product ward class assignment during the classification process is gen-
was obtained, there was no need to carry out any atmospheric erated through the mean decrease in Gini (MDG) coefficient.
correction operations. The RF classification was used to distinguish and effectively
characterize landcover impacted by oil pollution from oil-free
& Geometric correction vegetation. The analysis was carried out using the ImageRF
component of the EnMap Box (Waske et al. 2012). To achieve
In order to ensure that the Landsat 8 (OLI satellite image) this, various Vegetation Health Indices (generated in
co-registers properly with the other datasets (such as the oil BSampling regime^) together with seven Landsat (8 OLI
spill sites and boundary dataset), the satellite image was re- bands) (across visible, NIR, and SWIR) were fused for the
projected to the Universal Transverse Mercator projection and classification process. The tree size (ntree) used for classifica-
the World Geodetic Survey 1984 Datum of Zone 32 North tion was determined through repetitive runs before an optimal
(UTM WGS84 Zone 32N). value of 500 (ntree) was arrived at and used for parametriza-
tion in all classification scenarios implemented. Table 3 out-
& Landcover image masking lines the list of variables used for the RF classification.

Following the geometric correction of the study area image, Accuracy assessment
the three dominant existing landcover classes extracted from
the ESA CCI data (BLandcover data^) were used to subset the Two performance indicators were employed to assess the RF
image for the different landcover types. This provided the basis calibration model and the resulting classified image obtained.
of implementing a general study area wide classification oper- First is the F1 accuracy, which is the harmonic mean of pre-
ation (at macro level) and individual landcover subset classifi- cision and sensitivity (recall) accuracy statistics. This is used
cation (at micro level). The landcover image extent generated in the ImageRF to assess the out of bag error of the RF cali-
was for cropland, grassland, and TCA (i.e., dense canopy veg- bration. The precision is the ratio of correctly predicted posi-
etation), in which the harmonized oil spill and oil-free tive pixels to the total positive observations (incorporating true
landcover training sites were used to implement a macro and positives and false positives), while the recall is the ratio of
micro level classification. This produced six different landcover correctly predicted positive observations to the sum of true
schemes, that is, polluted (oil-impacted) cropland, polluted positives and false negative observations. This however can
grassland, polluted TCA, non-polluted (oil-free) cropland, be further interpreted as the measure of truly assigned pixels to
non-polluted grassland, and non-polluted TCA. a particular class (recall) and the measure of truly assigned
pixels in the image space. The F1 score is a robust accuracy
Retrieval of important Vegetation Health Indices measure for model performance. This is because it seeks to
balance the influence of recall and precision through the use of
Eight Vegetation Health Indices were generated using the for- harmonic mean of both measures.
mulae presented in Table 2. The indices were generated from This is denoted by the formulae below:
the pre-processed Landsat 8 (OLI image) of the study area
using the red, green, blue, near-infrared, shortwave infrared
1, and shortwave infrared 2 bands. Precision  Recall
F1 Accuracy ¼ 2  ð1Þ
Recall þ Precision
Random forest classifier TP
Precision ¼ ð2Þ
TP þ FP
The random forest (RF) algorithm was proposed by Breiman TP
(2001). It is an ensemble method for supervised classification Recall ¼ ð3Þ
TP þ FN
and regression, based on classification and regression trees
(CART). It relies on the assumption that different independent where
samples can influence positive predictions in different areas, TP = true positives
thus combining these true positives can significantly improve FP = false positives
overall prediction accuracy (Polikar 2006). The method also FN = false negatives
3626 Environ Sci Pollut Res (2019) 26:3621–3635

Table 2 Vegetation Health Indices generated using the red, green, blue, NIR, and SWIR bands

Vegetation indices Formula Author

Difference Vegetation Index RNIR − RRED Tucker 1980

 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Modified Soil-Adjusted Vegetation Index 1=2 2RNIR þ 1− ð2RNIR þ 1Þ−8ðRNIR −RRED Þ Qi et al. 1994
Moisture Stress Index RMidIR/RNIR Doraiswamy and Thompson 1982
Normalized Difference Vegetation Index (RNIR − RRED)/(RNIR + RRED ) Rouse Jr et al. 1974
Normalized Differential Water Index (RNIR − RSWIR)/(RNIR + RSWIR) Hardisky et al. 1983
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Renormalized Difference Vegetation Index RNIR −RRED = RNIR þ RRED Roujean and Breon 1995
Ratio Vegetation Index RRED/RNIR Jordan 1969
Soil and Atmospherically Resistant Vegetation Index (1 + 0.5) (RNIR − RRB)/(RNIR + RRB + 0.5) Qi et al. 1994
Soil-Adjusted Vegetation Index (1 + L)(RNIR − RRED)/(RNIR + RRED + L) Huete 1988
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Transformed Difference Vegetation Index RNIR −RNIR =ððRNIR þ RRED Þ + 0.5) Bannari et al. 2002

The error matrix as described by (Congalton 1991) was also accuracy, compared to the result obtained from the full image
used to assess the classified image output from the RF classifi- calibration. This shows that of the six schemes calibrated, the
cation using the 40% validation points (BSampling regime^). non-polluted (NP) and polluted (P) TCA and grassland re-
This enabled an effective comparison of the classified image spectively had better calibration result ranging from 45 to
outputs to the original reference sites. Specific attention was 70% F1 accuracy. While on the contrary, both the P and NP
given to the users, producers and the overall accuracies. croplands had lower calibration accuracies when the full study
area image was calibrated. The model calibration result also
showed that of the six different schemes investigated, the NP
Results grassland and NP TCA had the best prediction to error ratio of
86% and 84% as indicated in the F1 accuracy when the re-
RF model calibration spective landcover subsets were used. In contrast, the P and
NP croplands had the least calibration accuracy. In terms of
Figure 2 shows the result of the RF out of bag error. In general, the implication for interclass separability and model fit, it is
the result indicates that the landcover subset images had lower observed that calibration accuracy increased gradually from
out of bag errors and consequently higher calibration zero and mostly attained saturation when the tree size (ntree)
in the RF reached 50 using the variables, although for some
cases the F1 accuracy increased up to 100 trees before maxi-
Table 3 List of variables used for the RF classification
mum saturation was reached. This however implied that a
S/no Spectral variables lower ntree value could yield sufficient calibration result.
1 Band 1—ultra-blue band
2 Band 2—blue Landcover subset vs full image classification
3 Band 3—green
4 Band 4—red Figures 3 and 4 show the images classified from the two
5 Band 5—near-infrared (NIR) scenarios. The image classification at the landcover subset
6 Band 6—shortwave infrared (SWIR) 1 level had better representation of landcover extents with a
7 Band 7—shortwave infrared (SWIR)2 more generalized boundary compared to the full image clas-
8 Difference Vegetation Index (DVI) sification which had a crisper and noisy representation. This
9 Modified Soil-Adjusted Vegetation Index (MSAVI) however supports various assertions in several studies where
10 Moisture Stress Index (MSI) subpixel classification has been implemented (Aplin and
11 Normalized Differential Vegetation Index (NDVI) Atkinson 2001; Arif et al. 2015; MacLachlan et al. 2017). A
12 Normalized Differential Water Difference (NDWI) major reason for the observed disparity could be as a result of
13 Renormalized Difference Vegetation Index (RDVI) the presence of multiple signatures from conflicting landcover
14 Ratio Vegetation Index (RVI) features causing high spectral mixing for the RF classifier at
15 Soil and Atmospherically Resistant Vegetation Index (SARVI) the macro level. Fröhlich et al. (2013) have also observed that
16 Soil-Adjusted Vegetation Index (SAVI) textural characteristics of neighboring adjacent features can
17 Transformed Normalized Difference Vegetation Index (TNDVI) inadvertently cause false representation of image features.
Similarly, the spectral diversity of the features investigated
Environ Sci Pollut Res (2019) 26:3621–3635 3627

Fig. 2 RF parameterization result for the full study area image and F1 accuracy for the individual landcover subset images, while the red line
individual landcover subset images using training samples of oil-free represents F1 accuracy for the full study area image
and oil-impacted landcover. The green line represents parameterization

(polluted and non-polluted landcovers) had smaller separabil- schemes. This conforms with results obtained in Kalubarme
ity index as observed from the out of bag error for the full and Sharma (2015) where NDWI values were observed to be
study area image. This can affect the performance of the clas- sensitive to stress conditions in wheat-cultivated farm planta-
sifier in adequately producing generalizable extents. The im- tions. Similarly, results obtained by Benabdelouahab et al.
plication of this effect was further assessed using error matri- (2015) also showed that MSI and NDWI are sensitive indica-
ces generated. tors of stress also in a wheat-cultivated farm field. However,
the near-infrared and shortwave infrared bands were also ob-
Variable importance served to have the highest contribution in splitting oil contam-
inated and oil-free grassland landcover scheme. While the
The near infra band had the highest contribution to the assign- Difference Vegetation Index (DVI) and Normalized
ment of endmember classes for the six landcover schemes Differential Water Index clearly had strong contribution in
when the full study area image was classified (Fig. 5). Other splitting oil polluted from oil-free TCA.
variables however such as Moisture Stress Index, Normalized In general, the moisture-related indices and sensitive bands
Difference Water Index, shortwave infrared 1 (mid infrared (shortwave infrared 1) were observed to have more significant
region), and the green band also contributed substantially in contribution in distinguishing oil polluted from oil-free
the classification process. At the subset level, the result landcover types both at the macro level of the entire study
showed that the Normalized Difference Water Index and area and at the micro level of the individual landcover subsets.
Moisture Stress Index were very influential in providing the This is expected as the fundamental characteristics of stressed
best splits between polluted and oil-free cropland landcover vegetation are their inability to carry out basic life-supporting
3628 Environ Sci Pollut Res (2019) 26:3621–3635

Fig. 3 RF Image classification result for the full study area image and individual landcover subsets. It is observed that the former produced a more
generalized representation of landcover extents compared to the crisp output from the full study area image

Fig. 4 Subset of the study area showing the RF-classified image with the showed that spill-impacted and oil-free croplands were better captured
landcover subset of cropland into polluted and oil-free croplands. Inset is by the image subset classification (left), compared to the more crisp extent
a high-resolution image from Google Earth for the same area. This from the full study area image classification (right)
Environ Sci Pollut Res (2019) 26:3621–3635 3629

Fig. 5 Variable importance plot of the RF classification for the full study area image and landcover subset image classification

functions such as respiration, transpiration, and photosynthe- operation further reinforces the potentials of moisture-based
sis (Arellano et al. 2015), which the classifier can rely on from indices in depicting stress on vegetation. This plot showed the
the distinctions provided by the indices for class assignment. degree of variation in the health status of the oil-impacted and
Figure 6 shows the most important variables (i.e., NDWI, oil-free landcover classes. Non-polluted TCA were observed
SWIR, and DVI) in the classification process for cropland, to have the highest NDWI compared to the non-polluted crop-
grassland, and TCA landcover subsets respectively and their land and grassland. Generally, polluted grassland and crop-
respective oil-free and oil polluted landcover extents. land had the least NDWI greenness compared to their respec-
This shows that areas with high Vegetation Health Indices tive non-polluted classes. This is an indication that their health
and greenness are predominantly associated with oil-free status could have been affected by the oil spill in those loca-
landcover types especially for the oil-free cropland and grass- tions thereby accounting for lower health indices compared to
land landcover. While areas with low vegetation health and the respective oil-free vegetation. Similarly, the distribution of
greenness are mostly associated with polluted landcover the indices for the six classes shows little to no overlap be-
schemes in this case the polluted cropland and grassland. tween oil polluted and oil-free landcovers, a trend which could
However, TCA was noticed to have a poor split as indicative have accounted for the high performance of the NDWI in the
of the most important variable in the RF classification (Fig. 6). classification process.
This could be associated to the fact that large parts of the Niger
Delta are characterized by dense and mangrove forest vegeta-
tion (James et al. 2007), in which case the impact of crude oil
Accuracy assessment
would pose minimal discernible effect with a typical oil-free
The confusion matrix generated was used to evaluate the re-
vegetation.
sult of the RF classification for the two scenarios implemented
using the validation data (BOil spill incident data^) (Table 4).
Vegetation greenness distribution The overall accuracy from the full image classification gave
much lower accuracy (30.147%) compared to the result re-
Figure 7 is a box plot showing vegetation greenness retrieved corded from the various landcover subsets. Result from the
from NDWI for the various polluted and oil-free landcover tree cover densely forested areas gave the highest result of
training sites. This was the most influential index when the 70%, while the grassland and cropland subsets gave accura-
full study area image was classified together with the Moisture cies of 65% and 60.61% overall classification accuracy re-
Stress Index (MSI). Their performance in the classification spectively. In terms of interclass accuracy, the result from
3630 Environ Sci Pollut Res (2019) 26:3621–3635

Fig. 6 The most important variable for cropland—NDWI, grassland— the best split into oil-impacted and oil-free vegetation, as opposed to TCA
SWIR, and TCA—DVI in the classification processes. Result shows that subset where the most important variable did not give favorable split into
the most important variable for cropland and grassland classification had oil-impacted and oil-free TCA

Fig. 7 Box plot of vegetation

greenness retrieved from NDWI
for the polluted and oil-free
landcover samples
Environ Sci Pollut Res (2019) 26:3621–3635 3631

Table 4 Accuracy assessment result for the full study area and landcover-masked image classification

Map class Full image classification Masked classification

User’s accuracy Producer’s accuracy Overall accuracy User’s accuracy Producer’s accuracy Overall accuracy
[%] [%] [%] [%] [%] [%]

Non-polluted cropland 25 18.75 30.14 58.82 62.5 60.61

Polluted cropland 29.17 41.18 62.5 58.82
Non-polluted grassland 18.18 20 61.54 80 65
Polluted grassland 30 30 71.43 50
Non-polluted tree cover areas 50 40 75 60 70
Polluted tree cover areas 37.5 30 66.67 80

the validation exercise showed that the highest user accuracies Discussion
were obtained from the non-polluted grassland and polluted
TCA with 80% from the subset classification. Similarly, the Oil pollution and contamination of vegetation canopies within
landcover classes with the highest accuracy when the full the Niger Delta region is a common and almost a consistent
study area image was classified are the polluted and oil-free phenomenon. Few studies have focused on leveraging on the
TCA classes with producer and user accuracies of 50% and potentials of machine learning (ML) approaches (such as RF) to
40% respectively. This is not surprising as result from the map the exact oil spill extent for different landcover types. This
parameterization operation in Fig. 2 showed that the training study attempted to bridge this gap by using RF classification to
sites used for classification had better characterization be- first establish the precise extent of oil spill–impacted and oil-
tween polluted and oil-free dense canopies. Furthermore, the free landcover types. Then, secondly to identify the most useful
validation result obtained also showed that most of the classes optical indicators and discriminators of oil-impacted vegetation
that had better calibration also recorded higher accuracy. An communities from their respective oil-free vegetation. The re-
example is in the case of TCA and grassland schemes which sult obtained from these experiments after calibration of sample
recorded high accuracies of above 80% out of bag error, also sites and implementation of the classification operations
came out with 70% and 65% overall accuracies. showed that RF algorithm has the potential of providing reli-
able maps of oil-free and oil-impacted landcover. The RF clas-
Spill-impacted vs non spill landcover spatial extent sifier produced better results with the different landcover sub-
sets as opposed to when the full study area image is classified,
Figure 8a and b presents a stacked bar plot comparing the reinforcing the findings of Arroyo et al. (2010) where image
total estimated area covered by oil-impacted and oil-free space delineation for automatic classification of landcover fea-
landcover classes from the full study area and landcover tures proved very successful.
subset classification respectively. This was also compared The high calibration results obtained from the out of bag
to the total area coverage of the landcover product provid- errors during the parameterization exercise of the RF at the
ed by the ECCI. Generally, the result showed that aggre- micro level clearly account for the high accuracies of 70%
gated areas of polluted and oil-free landcover classes were and 65% obtained for the TCA and grassland vegetation types
closer to the areas from the ECCI when the image subsets respectively. Although the result of the most important variable
are classified than when the full image is classified. in the classification process (Fig. 6) does not mirror an excellent
Similarly, the extent of spill-impacted grassland and TCA split as can be observed with TCA and grass landcover subsets.
were larger than their respective oil-free vegetation, except A major reason for this trend can be attributed to the fact that
in cropland landcover where the area covered by oil-free most cropland vegetations are distinctly sparse in nature and a
cropland was larger than the oil-impacted cropland. In ad- huge volume of the oil spilt in these areas experience significant
dition, of the six landcover classes investigated the spatial seepage into the soil sub surface and immediately causing de-
extent of oil-impacted cropland from the full study area tectable impact on crops. This invariably accounts for a better
image and cropland landcover subset image classification split of oil polluted and oil-free croplands, as indicative of the
remained close. This, however, suggests that the spectral NDWI. Similarly, the exposed soil in cropland fields also
characteristics of the polluted cropland have remained un- means that much of the oiled sand surface reflective, account-
changed in the two experimental classifications imple- ing for the significant influence of shortwave infrared band
mented. This is an indication that this class could have (Ben-Dor et al. 1997; Cloutis 1989; Kühn et al. 2004) and its
been more heavily impacted from the 2015 and 2016 spill derived indices in distinguishing oil-impacted and oil-free crop-
incidents in the area. lands (Adamu et al. 2015; Ben-Dor et al. 1997; Brekke and
3632 Environ Sci Pollut Res (2019) 26:3621–3635

Fig. 8 Spatial extent of oil-

impacted and oil-free landcover
classes retrieved from the a full
study area image and b landcover
subset image classification. The
orange and blue stacks represent
polluted and oil-free landcover
classes, while the ash-colored line
represents the aggregate obtained
from the ECCI landcover dataset

Solberg 2005; Khanna et al. 2013; Kühn et al. 2004). This very suggests that cropland had the most significant impact, as
much infers that biomass density could play a significant role in the areas recorded from the full study area image and cropland
the characterization and mapping of oil polluted and oil-free landcover subset remained similar. This is quite contrary to the
terrestrial landcovers. results obtained from the TCA and grassland landcover, where
The variable importance plot obtained from the RF image the spatial extent of their polluted landcover had a much
analysis also showed that the near-infrared, shortwave infrared higher area than their non-polluted/oil-free landcovers. A pos-
bands, Normalized Difference Water Index, DVI, and MSI are sible reason for this trend could be as a result of over gener-
particularly influential in pixel class assignment. Some of these alization of the extent of spill-impacted landcover overlapping
variables (shortwave infrared, MSI, and NDWI) are mostly with other areas where vegetation stresses by other stressors
sensitive to vegetation moisture content (Gao 1996). Several exist. A post classification ground truth exercise carried out
studies (Agapiou et al. 2012; Arellano et al. 2015; showed that features such as waterlogged areas, dried vegeta-
Benabdelouahab et al. 2015; Dotzler et al. 2015; Kalubarme tion, burned vegetation, and cleared/exposed surface often
and Sharma 2015) have also shown that SWIR, MSI, and exhibited similar spectral signatures as polluted sites and were
NDWI variables are useful indicators of stress in vegetation classified as such. This is in line with observations made by
canopy as a result of their sensitivity to water net loss or gain. Khanna et al. (2013) and Kokaly et al. (2013). Although most
Similarly, the NIR band is also well known for its ability to of the aforementioned misclassification anomalies are also
distinguish between stressed and stress-free vegetations. This is vegetation stress related, accounting for the superior perfor-
because a major characteristic of a stress-free vegetation will be mance of the NDWI, NDVI, SWIR, and NIR in the classifi-
the absorption of visible light for photosynthesis necessary to cations processes. Figure 9 shows some the areas that exhib-
propagate the high reflectance of near-infrared energy (Ben- ited similar spectral response.
Dor et al. 1997; Knipling 1970). It is without doubt that these The problem of pixel misclassification in image classifica-
variables have the most ideal spectral information to character- tion is a general problem as also observed in (Ishida et al. 2018;
ize oil-free from oil polluted vegetation. The complex interac- Xiao and McPherson 2005; Zlinszky et al. 2012) where the
tion of these variables is a major reason for their incorporation characterization of a single vegetation type into a more narrow
in the classification process basically suggesting that stress as a group by species delineation or health status has been imple-
result of oil pollution can be better characterized and mapped. mented. The occurrence of pixel mismatch and over generali-
In addition, the result obtained from the spatial extent of the zation of landcover spatial extent is very much apparent in this
classified maps for polluted and oil-free landcovers further study. One way of addressing this problem in the future study is
Environ Sci Pollut Res (2019) 26:3621–3635 3633

Fig. 9 Potential influences to

pixel misclassification of an oil
polluted site. a Waterlogged
areas. b Cleared and exposed
surfaces. c Dried vegetated areas
which often lead to burning/burn
scars

the incorporation of other relevant variables (such as radar grassland and TCA, which can be addressed by the incorpora-
datasets, digital elevation model, soil-type map and soil mois- tion of other relevant variables in the classifier. In addition, the
ture) which generally do not specifically rely on the biochem- result of the variable importance showed that shortwave infra-
ical components of vegetation, rather the structural characteris- red and NDWI are significant variables in distinguishing oil
tics of vegetation and environmental factors are depended on to polluted and oil-free landcover, especially in cropland areas.
further improve discrimination accuracy. However, of the three oil polluted landcovers investigated, it
However, the concentration and size of spill also plays a is apparent that polluted cropland could have had the most
significant role in the detection and mapping of affected areas significant impact due to the similar result obtained (in terms
using the satellite image. Studies such as Adamu et al. (2016) of spatial extent) from the full study area and cropland image
have shown that the size of oil spill with respect to volume and subset classification. Similarly, the high distinctive split obtain-
age of oil is a major determinant of detectability of spill effect. ed from the NDWI (i.e., the most important RF variable) be-
This is largely predicated on the fact that not all spill incidents tween the oil-free and oil-impacted cropland areas, compared to
come in large sizes or quantities that can be meaningfully the TCA and grassland, is an indication of prolonged impact of
captured by the satellite sensors or pose detectable stress on hydrocarbon crude oil on the fragile cropland vegetation.
vegetation communities. In this study, we addressed this chal- The result obtained from this study certainly informs on the
lenge by using only spills with 1000 sqm or above in size to capability of using earth observation satellite data in charac-
ensure that the characteristics of a typical spill site are reason- terizing oil spill–impacted from oil-free areas even after sev-
ably captured within the spill epicenter and adjacent pixel eral months of spill occurrence. The successful application of
used for classification. It was however observed that other this method and approach to distinguishing these areas cer-
stress factors and features with same spectral characteristics tainly reinforces the potential of assessing the intrinsic linkage
can be potentially misclassified as oil polluted landcover, between oil-induced impacts and the concomitant long-term
which also transcend the results of the two image classifica- landcover changes. This will in no doubt provide a better
tion levels (micro and macro level) implemented. These cer- medium for assessing landcover change with specific recourse
tainly call for further research, especially using fuzzy tech- to oil spill incident in a typical oil spill prone area like the
niques in establishing precise spill threshold values for ade- Niger Delta region of Nigeria. Other limitations encountered
quate detection and classification purpose. in this study such as the lack of extensive cloud-free multi-
temporal optical images to establish phenological changes and
implement multi-temporal based classification can be system-
Conclusions atically addressed in future studies by incorporating radar
backscatter such as the freely accessible sentinel 1 SAR im-
This study aimed at applying RF in discriminating Landsat 8 ages in fostering the derivation of precise area extent of the
image pixels of oil polluted and oil-free landcover types using damage posed by oil pollution.
published oil spill incident records as the basis for formulating
training and validation sites. In addition, relevant Vegetation Acknowledgments Mohammed Shuaibu Ozigis was supported via a
Health Indices and image spectral bands were fused and classi- scholarship from the Petroleum Technology Development Fund (PTDF)
and National Space Research and Development Agency (NASRDA),
fied with RF classifier to support the discrimination process.
Nigeria. We also like to acknowledge the National Oil Spill Detection
Classification operation was implemented at the full study area and Response Agency (NOSDRA) and Shell Petroleum Development
(macro) level and at the individual landcover subset (micro) Corporation (SPDC) for making the oil spill incident record.
level. Results obtained from the latter gave a better characteri-
zation of oil-free from oil polluted landcover classes, as this Author Contributions This paper is the result of research conducted by
Mohammed Shuaibu Ozigis as part of his PhD studies under the supervi-
produced a more generalized extent compared to the crisp and sion of Jorg Kaduk and Claire Jarvis. Both Jorg Kaduk and Claire Jarvis
granular outputs produced from the former. Over generalization provided guidance in the technical design and implementation of the study,
and over estimation of the oil-impacted site were observed for as well as refinement of the initial manuscript to generate the final copy.
3634 Environ Sci Pollut Res (2019) 26:3621–3635

Compliance with ethical standards Cloutis E (1989) Spectral reflectance properties of hydrocarbons: remote-
sensing implications. Science 245(4914):1657168
Congalton RG (1991) A review of assessing the accuracy of classifica-
Conflict of interest The authors declare that they have no conflict of
tions of remotely sensed data. Remote Sens Environ 37(1):35–46
interest.
Doraiswamy PC, Thompson D (1982) A crop Moisture Stress Index for
Open Access This article is distributed under the terms of the Creative large areas and its application in the prediction of spring wheat
Commons Attribution 4.0 International License (http:// phenology. Agric Meteorol 27(1-2):1–15
creativecommons.org/licenses/by/4.0/), which permits unrestricted use, Dotzler S, Hill J, Buddenbaum H, Stoffels J (2015) The potential of
distribution, and reproduction in any medium, provided you give appro- EnMAP and Sentinel-2 data for detecting drought stress phenomena
priate credit to the original author(s) and the source, provide a link to the in deciduous forest communities. Remote Sens 7(10):14227–14258
Creative Commons license, and indicate if changes were made. Du S, Zhang F, Zhang X (2015) Semantic classification of urban build-
ings combining VHR image and GIS data: an improved random
forest approach. ISPRS J Photogramm Remote Sens 105:107–119
Fröhlich B, Bach E, Walde I, Hese S, Schmullius C, Denzler J (2013)
Land cover classification of satellite images using contextual infor-
References mation. ISPRS Annals of the Photogrammetry, Remote Sensing and
Spatial Information Sciences 3:W1
Gao B-C (1996) NDWI—A normalized difference water index for re-
Adamu B, Tansey K, Ogutu B (2015) Using vegetation spectral indices to mote sensing of vegetation liquid water from space. Remote Sens
detect oil pollution in the Niger Delta. Remote Sensing Letters 6(2): Environ 58(3):257–266
145–154. https://doi.org/10.1080/2150704x.2015.1015656 Giri C, Long J, Tieszen L (2011) Mapping and monitoring Louisiana’s man-
Adamu B, Tansey K, Ogutu B (2016) An investigation into the factors groves in the aftermath of the 2010 Gulf of Mexico oil spill. J Coast Res
influencing the detectability of oil spills using spectral indices in an 277:1059–1064. https://doi.org/10.2112/jcoastres-d-11-00028.1
oil-polluted environment. Int J Remote Sens 37(10):2338–2357. Gundlach ER, Hayes MO (1978) Vulnerability of coastal environments to
https://doi.org/10.1080/01431161.2016.1176271 oil spill impacts. Mar Technol Soc J 12(4):18–27
Adamu B, Tansey K, Ogutu B (2018) Remote sensing for detection and Hardisky M, Klemas V, Smart M (1983) The influence of soil salinity,
monitoring of vegetation affected by oil spills. Int J Remote Sens growth form, and leaf moisture on the spectral radiance of Spartina
39(11):3628–3645 alterniflora. Photogrammetric Engineering and Remote Sensing
Agapiou A, Hadjimitsis DG, Alexakis DD (2012) Evaluation of broad- 49(1):77–83
band and narrowband vegetation indices for the identification of Hester M, Mendelssohn I (2000) Long-term recovery of a Louisiana
archaeological crop marks. Remote Sens 4(12):3892–3919 brackish marsh plant community from oil-spill impact: vegetation
Alexakis DD, Sarris A, Kalaitzidis C, Papadopoulos N, Soupios P (2016) response and mitigating effects of marsh surface elevation. Mar
Integrated use of satellite remote sensing, GIS, and ground spectros- Environ Res 49(3):233–254
copy techniques for monitoring olive oil mill waste disposal areas on Huete AR (1988) A soil-adjusted vegetation index (SAVI). Remote Sens
the island of Crete, Greece. Int J Remote Sens 37(3):669–693 Environ 25(3):295–309
Aplin P, Atkinson PM (2001) Sub-pixel land cover mapping for per-field Ishida T, Kurihara J, Viray FA, Namuco SB, Paringit EC, Perez GJ, Marciano
classification. Int J Remote Sens 22(14):2853–2858 JJ (2018) A novel approach for vegetation classification using UAV-
Arellano P, Tansey K, Balzter H, Boyd DS (2015) Detecting the effects of based hyperspectral imaging. Comput Electron Agric 144:80–85
hydrocarbon pollution in the Amazon forest using hyperspectral Jackson JB, Cubit JD, Keller BD, Batista V, Burns K, Caffey HM,
satellite images. Environ Pollut 205:225–239. https://doi.org/10. Gonzalez C (1989) Ecological effects of a major oil spill on
1016/j.envpol.2015.05.041 Panamanian coastal marine communities. Science 243(4887):37–44
Arif M, Suresh M, Jain K, Dundhigal S (2015) Sub pixel classification of James GK, Adegoke JO, Saba E, Nwilo P, Akinyede J (2007) Satellite-
high resolution satellite imagery. Int J Comput Appl 129(1) based assessment of the extent and changes in the mangrove eco-
Arroyo, L. A., Johansen, K. and Phinn, S. (2010) Mapping land cover types system of the Niger Delta. Mar Geod 30(3):249–267
from very high spatial resolution imagery: automatic application of an Jhonnerie R, Siregar VP, Nababan B, Prasetyo LB, Wouthuyzen S (2015)
object based classification scheme, Proceedings of the GEOBIA. Random forest classification for mangrove land cover mapping
using Landsat 5 TM and ALOS PALSAR imageries. Procedia
Ayanlade A, Howard MT (2016) Environmental impacts of oil produc-
Environ Sci 24:215–221
tion in the Niger Delta: remote sensing and social survey examina-
Joel OF, Amajuoyi CA (2009) Physicochemical characteristics and mi-
tion. African Geographical Review 35(3):272–293
crobial quality of an oil polluted site in Gokana, Rivers State. J Appl
Bannari, A., Asalhi, H. and Teillet, P. (2002) Transformed difference Sci Environ Manag 13(3)
vegetation index (TDVI) for vegetation cover mapping, Jordan CF (1969) Derivation of leaf-area index from quality of light on
Geoscience and Remote Sensing Symposium, 2002. IGARSS’02. the forest floor. Ecology 50(4):663–666
2002 IEEE International. IEEE, pp. 3053-3055. Juel A, Groom GB, Svenning J-C, Ejrnæs R (2015) Spatial application of
Benabdelouahab T, Balaghi R, Hadria R, Lionboui H, Minet J, Tychon B Random Forest models for fine-scale coastal vegetation classifica-
(2015) Monitoring surface water content using visible and short- tion using object based analysis of aerial orthophoto and DEM data.
wave infrared SPOT-5 data of wheat plots in irrigated semi-arid Int J Appl Earth Obs Geoinf 42:106–114
regions. Int J Remote Sens 36(15):4018–4036 Kalubarme M, Sharma A (2015) Vegetation water stress assessment using
Ben-Dor E, Inbar Y, Chen Y (1997) The reflectance spectra of organic short wave infrared (swir) indices in wheat. Accessed Online Via
matter in the visible near-infrared and short wave infrared region Citeseerx. http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.
(400–2500 nm) during a controlled decomposition process. Remote 1.742.212&rep=rep1&type=pdf
Sens Environ 61(1):1–15 Khanna S, Santos MJ, Ustin SL, Koltunov A, Kokaly RF, Roberts DA
Breiman L (2001) Random forests. Mach Learn 45(1):5–32 (2013) Detection of salt marsh vegetation stress and recovery after
Brekke C, Solberg AHS (2005) Oil spill detection by satellite remote the Deepwater Horizon Oil Spill in Barataria Bay, Gulf of Mexico
sensing. Remote Sens Environ 95(1):1–13. https://doi.org/10.1016/ using AVIRIS data. PLoS One 8(11):e78989. https://doi.org/10.
j.rse.2004.11.015 1371/journal.pone.0078989
Environ Sci Pollut Res (2019) 26:3621–3635 3635

Knipling EB (1970) Physical and physiological basis for the reflectance Qi J, Chehbouni A, Huete A, Kerr Y, Sorooshian S (1994) A modified
of visible and near-infrared radiation from vegetation. Remote Sens soil adjusted vegetation index. Remote Sens Environ 48(2):119–126
Environ 1(3):155–159 Ramsey E 3rd, Meyer BM, Rangoonwala A, Overton E, Jones CE,
Kokaly RF, Couvillion BR, Holloway JM, Roberts DA, Ustin SL, Bannister T (2014) Oil source-fingerprinting in support of polari-
Peterson SH, Piazza SC (2013) Spectroscopic remote sensing of metric radar mapping of Macondo-252 oil in Gulf Coast marshes.
the distribution and persistence of oil from the Deepwater Horizon Mar Pollut Bull 89(1-2):85–95. https://doi.org/10.1016/j.marpolbul.
spill in Barataria Bay marshes. Remote Sens Environ 129:210–230. 2014.10.032
https://doi.org/10.1016/j.rse.2012.10.028 Ramsey E III, Rangoonwala A, Suzuoki Y, Jones CE (2011) Oil detection
Kuenzer C, van Beijma S, Gessner U, Dech S (2014) Land surface dy- in a coastal marsh with polarimetric synthetic aperture radar (SAR).
namics and environmental challenges of the Niger Delta, Africa: Remote Sens 3(12):2630–2662. https://doi.org/10.3390/rs3122630
remote sensing-based analyses spanning three decades (1986– Ramsey E, Rangoonwala A, Jones C (2015) Structural classification of
2013). Appl Geogr 53:354–368 marshes with polarimetric SAR highlighting the temporal mapping
Kühn F, Oppermann K, Hörig B (2004) Hydrocarbon index–an algorithm of marshes exposed to oil. Remote Sens 7(9):11295–11321. https://
for hyperspectral detection of hydrocarbons. Int J Remote Sens doi.org/10.3390/rs70911295
25(12):2467–2473 Roujean J-L, Breon F-M (1995) Estimating PAR absorbed by vegetation
Li M, Im J, Beier C (2013) Machine learning approaches for forest clas- from bidirectional reflectance measurements. Remote Sens Environ
sification and change analysis using multi-temporal Landsat TM 51(3):375–384
images over Huntington Wildlife Forest. GIScience & Remote Rouse Jr JW, Haas RH, Schell JA, Deering DW (1974) Monitoring veg-
Sensing 50(4):361–384 etation systems in the Great Plains with ERTS. Paper A 20, Remote
Liu M, Liu X, Li J, Ding C, Jiang J (2014) Evaluating total inorganic Sensing Center, Texas A&M University, College Station, Texa.
nitrogen in coastal waters through fusion of multi-temporal Accessed Via: NASA Technical Report Server
RADARSAT-2 and optical imagery using random forest algorithm. Roy DP, Wulder M, Loveland TR, Woodcock C, Allen R, Anderson M,
Int J Appl Earth Obs Geoinf 33:192–202. https://doi.org/10.1016/j. Kennedy R (2014) Landsat-8: science and product vision for terres-
jag.2014.05.009 trial global change research. Remote Sens Environ 145:154–172
MacLachlan A, Roberts G, Biggs E, Boruff B (2017) Subpixel land-cover Shi L, Zhang X, Seielstad G, Zhao C, He MX (2007) Oil spill detection
classification for improved urban area estimates using Landsat. Int J by MODIS images using fuzzy cluster and texture feature extrac-
Remote Sens 38(20):5763–5792 tion, OCEANS 2007-Europe IEEE, pp 1–5
Sun S, Hu C, Feng L, Swayze GA, Holmes J, Graettinger G, Leifer I
Mendelssohn IA, Andersen GL, Baltz DM, Caffey RH, Carman KR,
(2016) Oil slick morphology derived from AVIRIS measurements of
Fleeger JW, Overton EB (2012) Oil impacts on coastal wetlands:
the Deepwater Horizon oil spill: implications for spatial resolution
implications for the Mississippi River Delta ecosystem after the
requirements of remote sensors. Mar Pollut Bull 103(1-2):276–285.
Deepwater Horizon oil spill. BioScience 62(6):562–574
https://doi.org/10.1016/j.marpolbul.2015.12.003
Mishra DR, Cho HJ, Ghosh S, Fox A, Downs C, Merani PBT, Mishra S
Taheri, R. (2012) Oil slicks and coastal zones post Gulf War: a 20-years
(2012) Post-spill state of the marsh: remote estimation of the eco-
assessment, employing high-resolution satellite imagery, International
logical impact of the Gulf of Mexico oil spill on Louisiana Salt
Conference on Health, Safety and Environment in Oil and Gas
Marshes. Remote Sens Environ 118:176–185. https://doi.org/10.
Exploration and Production. Society of Petroleum Engineers.
1016/j.rse.2011.11.007
Tsutsumida N, Comber A, Barrett K, Saizen I, Rustiadi E (2016) Sub-
Noomen M, Hakkarainen A, van der Meijde M, van der Werff H (2015)
pixel classification of MODIS EVI for annual mappings of imper-
Evaluating the feasibility of multitemporal hyperspectral remote
vious surface areas. Remote Sens 8(2):143
sensing for monitoring bioremediation. Int J Appl Earth Obs
Tucker CJ (1980) A spectral method for determining the percentage of
Geoinf 34:217–225. https://doi.org/10.1016/j.jag.2014.08.016
green herbage material in clipped samples. Remote Sens Environ
Nriagu JO (2011) Oil industry and the health of communities in the Niger 9(2):175–181
Delta of Nigeria, Encyclopedia of Environmental Health. Elsevier UNEP (2011) Environmental assessment of Ogoniland.in UNEP Nairobi.
2011:240–250 Waske B, van der Linden S, Oldenburg C, Jakimow B, Rabe A, Hostert P
Obida CB, Blackburn GA, Whyatt JD, Semple KT (2018) Quantifying (2012) imageRF – a user-oriented implementation for remote sens-
the exposure of humans and the environment to oil pollution in the ing image analysis with Random Forests. Environ Model Softw 35:
Niger Delta using advanced geostatistical techniques. Environ Int 192–193. https://doi.org/10.1016/j.envsoft.2012.01.014
111:32–42 Whanda S, Adekola O, Adamu B, Pandey P, Ogwu F, Yahaya S (2016)
Ochege FU, George RT, Dike EC, Okpala-Okaka C (2017) Geospatial Geo-spatial analysis of oil spill distribution and susceptibility in the
assessment of vegetation status in Sagbama oilfield environment in Niger Delta region of Nigeria. J Geogr Inf Syst 8:438–456
the Niger Delta region, Nigeria. The Egyptian Journal of Remote Xiao Q, McPherson EG (2005) Tree health mapping with multispectral
Sensing and Space Science 20(2):211–221 remote sensing data at UC Davis, California. Urban Ecosystems
Onwurah I, Ogugua V, Onyike N, Ochonogor A, Otitoju O (2007) Crude 8(3-4):349–361
oil spills in the environment, effects and some innovative clean-up Zabbey N, Uyi H (2014) Community responses of intertidal soft-bottom
biotechnologies. International Journal of Environmental Research macrozoobenthos to oil pollution in a tropical mangrove ecosystem,
1(4):307–320 Niger Delta, Nigeria. Mar Pollut Bull 82(1-2):167–174. https://doi.
Park, Y. S., Al-Qublan, H., Lee, E. and Egilmez, G. (2016) Interactive org/10.1016/j.marpolbul.2014.03.002
spatiotemporal analysis of oil spills using comap in North Dakota, Zhu L, Zhao X, Lai L, Wang J, Jiang L, Ding J, Rimmington GM (2013)
Informatics. Multidisciplinary Digital Publishing Institute, p. 4. Soil TPH concentration estimation using vegetation indices in an oil
Pezeshki S, Hester M, Lin Q, Nyman J (2000) The effects of oil spill and polluted area of eastern China. PLoS One 8(1):e54028. https://doi.
clean-up on dominant US Gulf coast marsh macrophytes: a review. org/10.1371/journal.pone.0054028
Environ Pollut 108(2):129–139 Zlinszky A, Mücke W, Lehner H, Briese C, Pfeifer N (2012) Categorizing
Polikar R (2006) Ensemble based systems in decision making. IEEE wetland vegetation by airborne laser scanning on Lake Balaton and
Circuits and systems magazine 6(3):21–45 Kis-Balaton, Hungary. Remote Sens 4(6):1617–1650