Through the recent technological developments of radar and optical remote sensing in (i) the areas of temporal, spectral, spatial, and global coverage; (ii) the availability of such images either at a low cost or free of charge; and (iii) the advancement of tools developed in image analysis techniques and GIS for spatial data analysis, there is a vast potential for landslide studies using remote sensing and GIS as tools. Hence, this study aimed to assess the efficacy of using radar-derived factors (RDFs) in identifying landslide susceptibility using the bivariate information value method (InfoVal method) and the multivariate multi-criteria decision analysis based on the analytic hierarchy process statistical analysis. Using identified landslide causative factors, four landslide prediction models – bivariate with and without RDFs as well as multivariate with and without RDFs – were generated. Twelve factors such as topographical, hydrological, geological, land cover and soil plus three RDFs are considered. The weight of index for landslide susceptibility is calculated by using the landslide failure map, and susceptibility regions are categorized into four classes as very low, low, moderate, and high susceptibility to landslides. With the integration of RDFs, boundary detection between high- and very-low-susceptibility regions are increased by 7 % and 4 % respectively.
Landslides are one of the major types of geo-hazards in the world as approximately 9 % of global natural disasters are recorded as landslides (Chae et al., 2017; Chalkias et al., 2014). The recent statistics on landslide disasters per continent, from year 2000 to 2017, are summarized in the Emergency Events Database (EM-DAT, 2016). The database indicates that landslides cause around 16 500 deaths and affect 4.5 million people worldwide, with property damages of about USD 3.5 million (OFDA/CRED, 2016). The spatial prediction of landslide disasters, incorporating statistical analysis to identify areas that are susceptible to future land sliding, is one the important areas of geo-scientific research. These studies are based on the knowledge of past landslide events, topographical parameters, geological attributes, and other possible environmental factors (Park et al., 2013).
Presently, remote sensing has been used extensively to provide landslide-specific information for emergency managers and policy makers in terms of disaster management activities in the world (Baroň et al., 2014; Martha, 2011). The spatial resolution of space-borne optical data is now less than 1 m in panchromatic images, and at the same time synthetic aperture radar (SAR) sensors and related processing techniques have also increased. Radar is considered to be unique among the remote sensing systems, as it is all-weather, independent of the time of day, and is able to penetrate into the objects. Additionally, radar images have been shown to depend on several natural surface parameters such as the dielectric constant and surface roughness. The dielectric constant is highly dependent on soil moisture due to the large difference in dielectric constant between dry soil and water (Kseneman et al., 2012). It is accepted in the scientific community that remote sensing techniques do offer an additional tool for extracting information on the causes of landslides and their occurrences. Especially for deriving various parameters related to the landslide predisposing and triggering factors at global and regional scales, remote sensing plays a vital role (Corominas et al., 2014; Muthu et al., 2008; Pastonchi et al., 2018). Most importantly, landslide susceptibility analysis has greatly aided the prediction of future landslide occurrences, which is important for humans who reside in areas surrounded by unstable slopes. It is therefore identified that remote sensing techniques are significant in order to extract the landslide susceptibility regions by providing most suitable landslide predisposing factors at a smaller scale.
There is massive potential for research by applications in the area of disaster management if conventional remote sensing data and radar are integrated. This is because each method has its inherent disadvantages and shortcomings, as well as advantages, and integrating the two potentially complement each other. As such, this study combines the predisposing factors derived from both optical and radar satellite data for landslide susceptibility analysis. Furthermore, significant landslide predisposing factors like soil moisture, surface roughness, and forest biomass are derived from radar images, and the impacts of these factors on landslide susceptibility are examined.
There are inherent limitations and uncertainties in landslide susceptibility analysis, and yet several methods have been utilized and successfully applied in the past (Kanungo et al., 2009). These methods have been of both a qualitative and quantitative nature. Generally, qualitative methods are based on expert opinions while the quantitative approaches, such as statistical and probabilistic approaches, depend on the past landslide experiences.
Qualitative methods simply make use of landslide inventories to identify areas with similar geological and geomorphologic properties that show susceptibility to land failures. These methods can be divided into two groups: geomorphologic analysis and map combination. In geomorphologic analysis, the landslide susceptibility is determined directly either in the field or by the interpretation of images through geomorphologic analysis (Bui et al., 2011). Map combination is based on combining a number of predisposing factor maps for landslide susceptibility analysis. However, map combination analysis comprises a semi-quantitative nature by integrating the ranking and weighting of landslide susceptibility (Ayalew et al., 2004; Kavzoglu et al., 2014; Saaty, 1980). The analyses based on the quantitative approaches depend on numerical data and statistics, expressing the relationship between instability or predisposing factors with landslides (Reis et al., 2012). These methods are categorized into two groups: bivariate and multivariate statistical analysis. The popular information value method (InfoVal) is used as bivariate and multi-criteria decision analysis (MCDA) based on analytic hierarchy process (AHP) used as multivariate. Within the context of this work, these two methods are compared with respect to its performance in landslide susceptibility analysis.
It is understood that landslides may occur as consequences of complex predisposing and triggering factors. Topographical and geological factors, together with local climatic conditions, lead to landslide occurrences. The selection of these factors and preparation of corresponding thematic data layers are vital for models used in landslide susceptibility analysis (Jakob et al., 2006; Lee et al., 2017). There are no universal guidelines regarding the selection of predisposing factors in landslide susceptibility analysis. Some parameters may be important factors for landslide occurrences in a certain area but not for another one. Scientists (van Westen, 1997, 2003; van Westen and Getahun, 2003) show that every study area has its own particular set of predisposal factors which condition landslides. Determination of appropriate causal factors is a difficult task, and no specific rule exists to define how many factors are sufficient for a specific landslide susceptibility analysis. Hence, the selection of predisposing factors is dependent on the nature of the study area, opinions of the experts, and the availability of data for generating the appropriate spatial and thematic information (Kavzoglu et al., 2015; Shahabi and Hashim, 2015).
Koslanda in Sri Lanka is located at the geographical coordinates of
06
Topographical formation of Koslanda, Sri Lanka, with its previous landslide signatures. Sources © Google Earth and CNES.
The geomorphology of the area is described as a gently inclined talus slope,
with a thick, loosely compacted colluvium deposit at the foot of the near-vertical rocky scarp. Koslanda is situated at the middle part of the slope,
with the lower area showing a fairly steep surface as well. The composition
of the colluvium deposit in the area includes a randomly arranged mixture of
weathered clayey and sandy materials, with the organic matter making the
deposit act as a highly absorbing entity with high water content. The study
area was an abandoned tea land in which the properly maintained surface
drainage system has been neglected (Somaratne, 2016). Geology refers
to the physical structure and the substance of the Earth. The study area
consists mainly of undifferentiated charnockitic biotite gneisses and
quartzites, according to the
The most important phases in landslide prediction analyses are the collection of data from different sources and the construction of a spatial database on a common platform (Lan et al., 2004). The data utilized for the landslide susceptibility analysis include the topographical, hydrological, geological, soil, and land cover factors. All factors are derived from optical images (Landsat-8, Sentinel-2), radar images (Sentinel-1, TerraSAR-X), the digital elevation model (DEM) derived from aerial triangulation, and other available data sources (geology, rainfall). Stereo aerial photographs from 1993 are used to generate the 7 m resolution DEM using aerial triangulation (Copernicus Open Access Hub, 2018; USGS Earth Explorer, 2018). The landslide inventory map for the study area was constructed by integrating the interpreted multi-temporal aerial photographs, satellite images, and some temporal images from Google Earth (Fig. 2). Verifications are carried out through field investigations. In this research, the predisposing factors were selected from among the most widely considered factors in the literature and opinion from the experts.
Landslide catalogue of the Koslanda area with two different training and validating samples; background map from © Google Earth.
Most data are derived as primary data from remote sensing techniques for a large area with up-to-date information. As such, 15 predisposing factors are selected for the landslide susceptibility analysis by using bivariate and multivariate statistical techniques. Of these, 12 factors (elevation, slope, aspect, planar curvature, profile curvature, topographical wetness index (TWI), land use, lineament density, distance to water bodies, soil moisture, geology, and rainfall) are derived from optical images, DEM, and auxiliary data, while three more factors (soil moisture from delta index, surface roughness, and forest biomass) are derived from radar images. These factors were then combined in order to analyse the performance of this integration for landslide susceptibility analysis. Additional graphical information is available in Figs. S1, S2, and S3.
The topographical factors include elevation, slope, aspect, planar
curvature, profile curvature, and surface roughness of the terrain. The first
four factors are derived from the 7 m resolution DEM, and surface roughness
is derived using a Sentinel-1 radar image. The elevation is important to study
the local relief of the terrain and ranges from 446 to 1537 m above mean sea level in the
study area. Since the area contains high mountains, more than a 1000 m
difference in elevation can be observed. The basic parameter for the slope
stability analysis is the slope angle. The slope angle of the study area
ranges from 0
Under radar configuration, the magnitude of radar backscatter is defined as
a function of surface roughness and moisture content. Similar studies from
Rahman et al. (2008) and Septiadi and Nasution (2009)
emphasized the extraction of surface roughness from radar data using
textural analysis. Hence, to estimate the surface roughness without the use
of any ancillary field data, a Sentinel-1 radar image on 12 March 2015 under dry climatic conditions was used to reduce the effect of the
moisture component from the radar backscatter. The texture is the structure,
or appearance, of the surface and, as such, describes the coarseness or the
homogeneity of the image structure. One of the most prominent methods for
texture analysis is the grey-level co-occurrence matrix (GLCM), which is based
on the second-order probability density function. The GLCM describes how
often a grey level occurs at a pixel located at a fixed geometric position
relative to its neighbourhood pixels. The surface roughness is normally a
measure of finer surface irregularity in the surface texture. These texture
features extracted from the GLCM would be the best descriptors for
quantifying the state of surface roughness (Septiadi and
Nasution, 2009). Hence, the GLCM texture analysis is performed using a
window size of 9 pixels
Distance to hydrological features, rainfall, and TWI defined by Eq. (1) are
selected as the hydrological factors for this landslide susceptibility
analysis. Proximity to the hydrological features is an important factor when
considering the landslide susceptibility analyses (Sar et
al., 2016; Shahabi and Hashim, 2015). TWI is a solid index that is capable
of predicting areas susceptible to saturation or wetness of land surfaces,
as well as the areas that have the potential to produce an overland flow. Within
the Sri Lankan context, heavy and prolonged rainfall is the main triggering
factor for the landslides. The monthly average rainfall data for the years
2014 to 2016 from 10 nearby stations to Koslanda were used in this study.
Monthly rainfall data from 10 rain gauge stations are averaged, and the
average rainfall map for the study area is generated using the inverse
distance weighting (IDW) interpolation method within the ArcGIS environment.
TWI has been used to study the spatial scale effects, or topographic
control, on hydrological processes. This index was developed by Beven and
Kirkby (1979) and can be defined in Eq. (1) as
The soil moisture index (SMI) defined in Eq. (2) and the delta index defined in
Eq. (5) are the soil factors focused upon in this research. Surface soil
moisture is one of the most important parameters in land susceptibility
analysis (Carlson et al., 1994; Zhan et al., 2002). Several
methods have been proposed to estimate the surface soil moisture conditions
accurately with in situ measurements. However, these methods are time-consuming and costly when the area of interest is large and the scale of
work is small. Hence, this research uses the universal triangle relationship
between soil moisture, the normalized difference vegetation index (NDVI), and
land surface temperature (LST) derived from Landsat-8 image bands as an
optical remote sensing approach, as well as the delta index derived from two radar
images, as wet and dry conditions, as a radar remote sensing approach. Band
5 (near-infrared, NIR; 30 m resolution), band 4 (red, 30 m resolution), and
band 11 (thermal, TIR-2, 100 m resolution) of the Landsat-8 image of 3 July 2015 are processed for extracting the soil moisture index in the
thermal-NDVI space. The SMI is “0” along the dry edge and “1” along the wet
edge. According to the studies from Wang and Qu (2009) and Zenga et
al. (2004), SMI can be defined in Eq. (1) as
The major land uses existing in this study area are identified as tea, scrub, forest, rock, rice, water, and residential. The Sentinel-2A image from 10 October 2016 is used to extract the desired land uses from the study area by applying supervised classification. Scrub areas are typically the tea estates that are in abundance, while the residential areas are the rooms of tea workers. It is noted that most of the devastating landslides in this area had occurred within the extensive tea estates. Hence, the main reason for the continuous occurrence of these landslides can be identified as the lack of proper land use management in the area.
Forest biomass is a significant factor that can control the landmass
failures or landslides. The main limitations of using optical remote sensing
for forest biomass estimation are the near-constant tropical cloud cover and
the insensitivity of reflectance to change in the biomass in older and mixed
forests. Radar has the potential to overcome the above limitations due to its
all-weather, day and night capability, with the positive relationship of
radar backscatter and forest biomass. The spatial, spectral, temporal, and
polarization characteristics of radar backscatter have a known influence on
the forest biophysical properties. Kuplich et al. (2005) and Caicoya
et al. (2016) related the radar image texture derived from GLCM to
the forest biomass. An experiment was conducted by Kuplich et al. (2005) with seven texture measures, but only the GLCM-derived
contrast increased the correlation between the backscatter and the log of
biomass in Eq. (6) as
Geology refers to the physical structure and the substance of the Earth. In
order to investigate the land mass failures, the geological structure of
that particular area has to be analysed carefully. In addition to the
geology of the area, lineament density has also been considered as a factor.
The geological information of the particular area is obtained from the
geological map available at the Geological
Survey and Mines Bureau, Sri
Lanka, at a
The InfoVal method determines the susceptibility at each point or pixel, jointly considering the weight of influence of all predisposing factors. The weight of influence is based on the landslide inventory map of the particular area. When constructing a probability model for landslide prediction, it is necessary to assume that the landslide occurrence is determined by landslide-related factors and that future landslides will also occur under the same, or almost similar, conditions as past landslides (Remondo et al., 2013; Saha et al., 2005). Hence, at the beginning of the analysis, the landslide inventory map is divided in to two samples – training and validation – enabling the use of these data for landslide susceptibility analysis and validation of results respectively as in Fig. 2. The log function is used to control the large variation of weights in calculations. The larger the weight of influence, the stronger the relationship between landslide occurrence and the given factor's attribute.
This method overlays all individual predisposing factors such as thematic maps
with the landslide inventory map to calculate the density of the landslide
detachment zones for each class of the selected factors. The density of
landslide pixels represents the weight of influence of each predisposing
factor in Eq. (7) as
The MCDA method integrates all the independent predisposing factors with the inclusion of relative contribution of each factor by putting more emphasis on the predisposing factors that contribute to landslide occurrence. The same predisposing factors with or without radar are used to investigate the landslide susceptibility regions from the AHP technique within the GIS domain. In AHP, each pair of factors in a particular factor group is examined at one time, in terms of their relative importance. Relative weights for each factor are calculated based on a questionnaire survey from experts in the field (further information for the questionnaire is available in Fig. S4). These relative weights are then used to generate a pairwise comparison matrix, which is the basic measurement mode when applying the AHP procedure. The selected predisposing factors, and relevant relative weights, are used to generate the normalized matrix with final average weights. However, expert knowledge could be subjective at times, or may cause one to assign different weights for each factor, when dealing with a large number of causative factors. Hence, in order to avoid this inconsistency, the consistency ratio (CR) is calculated. For better predictive models, the CR should be less than 0.01, otherwise each factor has to be generated with the proper pairwise comparison.
The calculated final weights for 12 landslide predisposing factors
without RDFs such as elevation–slope, aspect, planar curvature, profile
curvature, TWI, land use, lineament density, distance to water bodies, soil
moisture, geology, and rainfall were 0.030, 0.172, 0.022, 0.018, 0.014,
0.074, 0.149, 0.052, 0.045, 0.094, 0.185, and 0.145, respectively. The CR is
0.089, making it less than 0.1, which is the value shown to be the reasonable level of consistency in the pairwise comparison. The final weights for 15
predisposing factors with RDFs – elevation, slope, aspect, planar curvature,
profile curvature, TWI, land use, lineament density, distance to water
bodies, SMI in NDVI-
After decisive analysis of the types of predisposing factors, the presented work proceeded to consider 15 predisposing factors that are derived from optical, radar, and other available auxiliary data sources. Three significant causative factors – surface roughness, soil moisture from the delta index, and forest biomass – were estimated by using radar satellite images. Thus, this work investigated the performance of landslide susceptibility analysis using bivariate and multivariate methods with the inclusion of RDFs and described the processing steps in Fig. 3.
Workflow of the landslide susceptibility analysis using bivariate and multivariate approaches.
The weight of influence of all predisposing factors such as thematic maps is added in a bivariate and multivariate manner to obtain the contribution of all predisposing factors for landslide susceptibility analysis. After calculating the cumulative percentage of failures of the weighted susceptibility maps, value ranges for each percentage of failure are obtained from quantile classification for 10 classes. The entire study area of each landslide susceptibility map is then discretized in to four classes as 0 %, 10 %, 30 %, and 60 % of failure regions for very-low-, low-, moderate-, and high-susceptibility classes, respectively.
Four landslide prediction models, (i) bivariate without RDFs (BiNR), (ii) bivariate with RDFs (BiWR), (iii) multivariate without RDFs (MNR), and (iv) multivariate with RDFs (MWR), are discussed. The region has been analysed and classified into four (04) landslide susceptibility regions: high, moderate, low, and very low.
Susceptible regions are identified from the bivariate InfoVal method without
RDFs as 12 % for high, 45 % for moderate, 38 % for low, and 5 % for
very low as shown in Fig. 4a. Hence, 57 % of areas from the total study area
are predicted to have high and moderate susceptibility to landslide
hazards. Very steep slope mountains in the north, north-west, and east
regions are identified as very-low-susceptibility areas, given that the area
was free from historical landslides. The middle regions with
30
Landslide susceptibility maps from bivariate and multivariate
analysis with and without RDFs.
All 15 weighted predisposing factors were grouped as with and without RDFs, and the weighted overlay is performed separately in order to obtain the landslide susceptibility regions.
Figure 4c illustrates the landslide susceptibility map from the
multivariate method without RDFs and is able to identify 18 % for high-,
44 % for moderate-, 36 % for low-, and 2 % for very-low-susceptibility
regions. Hence, 62 % of areas from the total study area are predicted to
be of high and moderate susceptibility to landslide hazards. In the
landslide susceptibility map from the multivariate method with RDFs, from the
total area, 21 % of the area shows a high susceptibility to landslides,
with 40 % of the area as moderate, 34 % of the area as low, and 5 % of the area as
having very low susceptibility as shown in Fig. 4d. Hence, 61 % of areas
from the study area are predicted to have high and moderate susceptibility to landslide hazards. In a similar manner to the InfoVal method, the top
of the mountains in the north, north-west, east, and south-east regions,
the area near to the Eruwendumpola Oya, are identified as having a very low
susceptibility to landslide hazards, while the middle regions with
30
Landslide prediction is of utmost importance in all phases of disaster management and development activities in a country. In recent years Koslanda in Sri Lanka has been found to be significantly prone to landslide disasters. Hence, this study investigated the efficacy of radar-derived factors for landslide susceptibility analysis of a bivariate and multivariate nature. The main difference between bivariate and multivariate analysis is that in multivariate analysis the predisposing factors are weighted by considering how each of them affect landslide hazard. Four landslide susceptibility maps are produced from bivariate and multivariate analysis with and without radar-derived factors. The areas identified as having high- and moderate-susceptibility classes in these four approaches (57 %, 58 %, 62 %, and 61 % respectively in BiNR, BiWR, MNR, and MWR) are close in value but show an increase in multivariate analysis when compared with bivariate analysis as tabulated in Table 1. Moderate- and low-landslide-susceptibility areas show very small ((1–2) %) changes between these four types of analysis. With the integration of RDFs such as surface roughness, near-surface soil moisture from the delta index, and forest biomass in bivariate and multivariate analysis, the high- and very-low-susceptibility areas are increased significantly (high: 7 % – bivariate, 3 % – multivariate; very low: 4 % – bivariate, 3 % – multivariate). However, when comparing the high- and very-low-susceptibility areas from bivariate and multivariate analysis, high-susceptibility areas show a considerable increase (without radar: 6 %; with radar: 2 %) while very-low-susceptibility areas have a noteworthy decrease (without radar: 3 %; with radar: 4 %).
Landslide-susceptible area comparison from bivariate and multivariate analysis with and without RDFs. BiNR – bivariate analysis without RDFs, BiWR – bivariate analysis with RDFs, MNR – multivariate analysis without RDFs, and MWR – multivariate analysis with RDFs.
The landslide susceptibility maps derived from the bivariate and multivariate analysis are validated using the selected validation samples from the landslide failure map. The most commonly used and scientifically recognized receiver operating characteristic (ROC) curves are used to analyse the prediction and validation performances. The ROC curve is a graphical plot that illustrates the performance of classification and is considered to be a powerful tool for the validation of landslide susceptibility analysis for many years (Neuhäuser et al., 2012). The areas under the curve (AUCs) for the four different approaches – bivariate and multivariate with and without RDFs – are calculated and graphed in Fig. 5.
Success rate and prediction rate curves with AUC for the bivariate
and multivariate analysis with and without RDFs. The
The areas under the success rate curves measure how the landslide prediction analysis fit with the training data set, while the areas under the prediction rate curves measure how well the landslide prediction models and landslide causative factors predict the landslides. If the area under the ROC curve is closer to 1, the result of the test is excellent and vice versa, and when AUC is closer to 0.5, the result of the test is fair or acceptable (Kamp et al., 2008).
The AUCs of all the success rates are more-or-less near 0.80, indicating good prediction performances according to the definition. The AUCs of all the prediction rates are having values above 0.50, thereby indicating that they are within the acceptable range as per the definition. As such, they indicate that the accuracy of the prediction rate of land susceptibility and the selection of land causative factors are acceptable, but not excellent, even though the fit between the landslide prediction and the training data set is excellent as compared in Table 2. The incompleteness of the available landslide inventory map, as well as an insufficient number of validation samples in the study area, can be shown as reasons for the discrepancy. As a whole, better prediction and validation capabilities are shown by the bivariate analysis when compared with the multivariate approaches.
Comparison of area under success rate and prediction rate curves for bivariate analysis without RDFs (BiNR), bivariate analysis with RDFs (BiWR), multivariate analysis without RDFs (MNR), and multivariate analysis without RDFs (MWR).
This study focused on the applicability of remote sensing and GIS for rapid landslide prediction analysis at a finer scale. Furthermore, by considering the significance of radar data for landslide analysis, this study mainly investigates the efficacy of radar-derived factors for landslide prediction analysis, which is not well experimented in the current research. Most significant predisposing factors such as surface roughness, soil moisture, and forest biomass derived from radar are incorporated to examine the landslide prediction analysis. The prediction analysis is performed by using bivariate and multivariate statistical analysis.
The main difference between bivariate and multivariate analysis is that in
multivariate analysis selected predisposing factors are weighted by
considering how each of them are influenced for landslide susceptibility.
This study investigated 15 landslide predisposing factors: elevation,
slope, aspect, planar curvature, profile curvature, TWI, land use, lineament
density, distance to hydrology, SMI in NDVI-
From the results obtained, it can be concluded that the bivariate and multivariate statistical analysis, with and without RDFs, can be used for landslide prediction analysis. However, with the integration of RDFs such as surface roughness, near-surface soil moisture from the delta index, and forest biomass, the detection of the boundary between the high- and very-low-susceptibility regions is increased. When comparing the bivariate analysis with multivariate, the increase of high- and very-low-susceptibility regions is higher in bivariate than multivariate. In landslide prediction analysis, the most important susceptibility classes are the high- and very-low-susceptibility classes, as they provide significant information about the danger from a disaster. Hence, with the integration of radar-derived factors, by increasing the accuracy of prediction for high-susceptibility regions, the possibility of mitigating dangers can be considerably improved. When the accuracy and prediction of very-low-susceptibility regions are increased, the use of such lands can be encouraged for residential, community places, and safe areas when a landslide occurs.
Successful prediction and validation of prediction analysis via ROC curves are achieved. Even though this study was tested for a sample area, the same methodology can be applied for any landslide-prone area to investigate the landslide prediction analysis using radar-derived factors by using bivariate and multivariate analysis. This is because the radar-derived factors can be derived for any area, as long as the data are available, and at any time under any weather conditions as radar is weather independent. Additionally, the technology can be learned easily and anyone can be trained to use this methodology to predict landslide susceptibility areas, and this is especially helpful for developing countries which do not have up-to-date data at fine resolutions. With the increasing availability of free data from optical sensors, radar sensors, and DEM, it is possible to derive more landslide predisposing factors such as thematic maps. Furthermore, there are many statistical analyses developed of a qualitative and quantitative nature for spatial data analysis. Hence, further investigations have to be performed for landslide susceptibility analysis, even focusing on the changing nature of the environments.
Landsat-8 data can be freely downloaded from USGS Earth Explorer (2018),
The supplement related to this article is available online at:
AKRNR performed the conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, validation, visualization, and writing of the original draft; RB supervised the study and reviewed and edited the original draft of the manuscript. UGAP and TLD supervised the study.
The authors declare that they have no conflict of interest.
The authors wish to acknowledge the Sabaragamuwa University of Sri Lanka for offering an opportunity for this research and the HETC project, Ministry of Higher Education, Sri Lanka, for providing financial support under the grant number SUSL/O-Geo/N2 as well as the University of Siegen, Germany, for providing their support in collecting and initial processing of the TerraSAR-X images from DLR, Germany. The DLR, Germany, is acknowledged with appreciation for providing radar images free of charge, and the GSMB, Sri Lanka, is acknowledged for providing the geological data necessary for this research work freely.
This research has been supported by the HETC project, Ministry of Higher Education, Sri Lanka (SUSL/O-Geo/N2).
This paper was edited by Filippo Catani and reviewed by three anonymous referees.