Combination of optical images and SAR images for detecting landslide scars, using a classification and regression tree

ABSTRACT Landslides are some of the most destructive and recurrent natural hazards worldwide. Landslides are triggered by natural phenomena such as extreme rainfall and earthquakes, causing human and economic losses. A rapid response to landslide events is necessary to assess damage mitigation and save lives and property. This study developed a landslide detection model using differential spectral indices and amplitude ratio changes with a classification and regression tree (CART), aiming to detect landslide scars after the occurrence of these events in Asian regions for testing different environment condition. The multi-temporal SAR and optical stack images were pre-processed to reduce speckle noise, seasonal noise, and atmospheric noise. This study explored change detection approaches with a minimum threshold of amplitude ratio change (Aratio), using Sentinel−1 images and the relative difference in the normalized difference vegetation index (rdNDVI), differential bare soil index (dBSI), and differential brightness index (dBI) was obtained using Sentinel−2 images. The accuracy of the model was examined by F1-scores. The accuracy of the model for landslide detection was considered moderately good to excellent. As a result of the landslide detection model, amplitude ratio change detection improved the model as revealed by the F1-scores. Moreover, this study found that differential spectral indices could be used to classify the types of landslides (deep-seated and shallow landslides) according to the level of surface changes and texture of the collapsed material after landslide events.


Introduction
A landslide is defined as the movement of a mass of earth down a slope, which is triggered by heavy rainfall from a typhoon, melting snow, earthquakes, and human activities like timber harvest and road construction (Aleotti and Chowdhury 1999). Between 2004 and 2016, landslides occurred globally, causing substantial damage to people, property, and economies. The total losses during this period exceeded 55,000 human lives and approximately USD 20 billion annually (Sim, Lee, and Wong 2022). Detecting landslides rapidly is crucial for creating accurate inventories in the weeks and months following a trigger event, such as heavy rainfall or an earthquake. These inventories serve as critical tools for reducing the risk of landslides by identifying and mapping areas of landslide susceptibility as well as regions that have experienced landslides in the past. Accurate landslide inventories are important for understanding the landslide mechanism and developing landslide susceptibility models for prevention and mitigation. Landslide inventories should be detailed and include information about the location, land cover, morphology, type of landslide, and trigger. These inventories improve our understanding of landslide occurrences and landslide types, can help quantify landslide damage and erosion area, and allow the observation of secondary hazards such as debris flows and landslide dams (Kirschbaum and Stanley 2018;Kirschbaum, Stanley, and Zhou 2015;Zhong et al. 2020). Furthermore, this information can be used to prioritize mitigation efforts, such as stabilizing slopes, building retaining walls, and planning for emergencies and mitigation. Recently, landslide inventories have been prepared by applying various techniques including aerial photography and optical image visualization, synthetic aperture radar (SAR) imagery, light detection and ranging (LiDAR), field surveys, and local community reports. Satellite remote sensing has proven to be useful for generating landslide inventories and assisting traditional time-consuming mapping methods that mostly rely on field surveys and visual interpretation of aerial photographs (Guzzetti et al. 2012).
Recently, earth observation (EO) imagery and related processing and visualization techniques and tools have notably increased the ability to prepare landslide maps (Mondini et al. 2021). Remote sensing techniques are typically used to construct landslide inventories over areas that have experienced catastrophic events (Bessette-Kirton et al. 2019;Roback et al. 2018). Several studies have used satellite-based optical images, such as those provided by SPOT−5, Landsat−8, and Sentinel−2, and aerial images with manual mapping approaches to provide high-quality information regarding the landslide (Amatya, Kirschbaum, and Stanley 2019;Massey et al. 2020;Shahabi and Hashim 2015). Recent landslides usually appear brighter than their immediate surroundings because of the exposure of the bare ground and they can be interpreted directly by visualization, facilitating the differentiation of pre-event and postevent images (Li et al. 2016;Nichol and Wong 2005). In addition, the spectral properties of optical images are mainly used with thresholding spectral indices for landslide recognition purposes (Behling et al. 2014;Hölbling et al. 2018;Scheip and Wegmann 2021). Previous studies on landslide detection have applied spectral index thresholds that are region-specific and, therefore, performed poorly when applied to new areas (Tehrani, Santinelli, and Herrera Herrera 2021). Moreover, the spectral indices have not been applied thoroughly to determine the landslide characteristics. To classify the landslide characteristics, geomorphological, geological, hydrological, and other factors are considered when applying the spectral indices (Dou et al. 2015;Zhong et al. 2020). Furthermore, the success of using optical images (both aerial and satellite) strongly relies on the weather at the time of acquisition. Optical images are limited in terms of a quick response because they require sunlight and cloud and shadow-free conditions to identify landslides accurately (Williams et al. 2018). Nevertheless, the satellitebased SAR can minimize these weather and time of acquisition limitations.
SAR can penetrate clouds and acquire data day and night. This can be beneficial for the rapid mapping of landslides triggered by intense or prolonged rainfall. However, its limitations include geometric shading, speckle noise, and distortion. SAR data have been used for landslide mapping and monitoring (Colesanti and Wasowski 2006;Ferretti, Prati, and Rocca 2001;Luo et al. 2016;Nishiguchi, Tsuchiya, and Imaizumi 2017;Ohki et al. 2020). Most studies have focused on SAR interferometry (InSAR), which primarily detects surface variations caused by landslides. Recently, differential interferometric synthetic aperture radar (D-InSAR) techniques and phase differences in SAR images have been used for detecting individual landslides or single events from ground displacement for slow-moving (creep) landslides with a multitemporal time series Righini, Pancioli, and Casagli 2012;Solari et al. 2018). The calculated vertical and horizontal displacement velocities from InSAR are employed to analyse the surface ground movement for loess soil landslides (translational slides, rotational slides, and loess flows) in the Huangshui region of China (Meng et al. 2020). Furthermore, the SAR radar amplitude and coherence-based detection of changes can also be used to identify landslides when there are changes in the properties of the ground surface (e.g. reflectance, roughness, or dielectric properties) of pre-and post-landslide events (Mondini et al. 2019). Coherence-based change-detection methods perform effectively in urban areas because coherence is generally high before a landslide event, and is reduced as a result of damage after the event. In mountainous areas with dense vegetation, the amplitude-based changedetection method outperforms the coherence-based method because of the consistently low coherence change. By contrast, the amplitude change is apparent in landslide areas in mountainous areas (Jung and Yun 2020). Mondini et al. (2017) used continuous measurements of SAR amplitude changes and spatial autocorrelation to intercept hundreds of landslides in Myanmar (Mondini et al. 2017). Konishi and Suga (2018) investigated the potential of the backscattering-coefficient difference and the intensity correlation between pre-and post-COSMO-SkyMed images for landslide detection in the Kii Peninsula (Konishi and Suga 2018). Mondini et al. (2019) used the SAR amplitude changes of pre-and post-Sentinel−1 images for landslide detection in 32 worldwide cases of rapid landslides entailing different types, sizes, slopes, and triggers. Although the detection of SAR amplitude changes in bitemporal images may detect and classify landslides, the accuracy of landslide detection is affected because of speckle noise and atmospheric noise. Thus, increasing the number of SAR images from pre-and post-event stacks could reduce the noise from radar scattering and speckle in the SAR data (Handwerger et al. 2020;Lindsay et al. 2022).
Using both optical and radar satellite images for landslide detection is rare and has been mentioned in only a handful of published studies. The first recorded research that applied optical and radar data identified the diagnostic characteristics of a landslide in Canada, using Landsat TM and RADARSAT data (Singhroy, Mattar, and Gray 1998) (Plank, Twele, and Martinis 2016), used very high-resolution polarimetric SAR obtained with Landsat 8 and Sentinel−2 for landslide detection (Plank, Twele, and Martinis 2016). Later (Mwaniki et al. 2017), used SAR data, normalized difference mid-infrared spectral index (NDMIDIR) from Landsat 8 optical imagery, and geologic features to map landslides in central Kenya (Mwaniki et al. 2017) (Hölbling et al. 2018), integrated optical images and SAR intensity data with morphological properties to distinguish the types of landslides (shallow landslide and debris flow). The SAR intensity from Sentinel−1 and the atmospherically resistant vegetation index (ARVI) from Sentinel−2 were utilized to detect landscape changes triggered by the 2016 Kaikoura earthquake in New Zealand (Jelének and Kopačková-Strnadová 2021). Moreover, the combination of optical images and SAR images was integrated with machine learning to detect landslides and map susceptibility in the Cameron Highlands, Malaysia (Bui et al. 2018;Nhu et al. 2020). D-InSAR by Sentinel−1 and optical image data used to apply for identification and monitor deformation, and the integration approach could apply landslide mitigation for slowdeformation landslide . SAR image data can exhibit low coherence in vegetated mountainous areas, and atmospheric and meteorological conditions frequently limit optical image data at the acquisition time. Therefore, SAR image data have been combined with optical images to minimize these limitations. Although some of the above-mentioned studies aimed to develop an integration of optical images and SAR images with semi-automated and machine-learning tools for landslide detection, they utilized bitemporal images from pre-and post-events and fared poorly in landslide detection due to noise from agriculture, the weather, and the atmosphere. Recently (Lindsay et al. 2022), examined Sentinel−1 and Sentinel−2 to compare the landslide visibility between bitemporal and multitemporal change detection of landslides on a glacial landscape in Norway. The multitemporal change detection outperformed the bitemporal one in terms of noise reduction. Therefore, this study tested multitemporal images to reduce the speckle, atmospheric, and meteorological noise. The advantages of SAR images over optical images include their ability to acquire data during both day and night and their ability to penetrate clouds. Moreover, minimum differential spectral index thresholds have been obtained by several study areas and used to classify the landslides, which is a novel approach. In addition, this study examined minimum thresholds for eight study areas in Asia regions and validate with the historical landslide inventory.
The main objective of this study was to investigate how a combination of multitemporal SAR and optical images can be used to detect landslides under different conditions and improve the accuracy of landslide detection. The classification and regression tree (CART) machine-learning approach was applied to evaluate the identification of landslides in the classification phase. This study explored change detection approaches with a minimum threshold of amplitude ratio change (A ratio ), using Sentinel−1 images and the relative difference in the normalized difference vegetation index (rdNDVI), differential bare soil index (dBSI), and differential brightness index (dBI) was obtained using Sentinel−2 images. This study also aimed to compare the differences among data sets and classify the landslides into shallow and deep-seated landslides by applying the results of landslide detection obtained using the images processed with both amplitude ratio and differential spectral indices thresholding.

Study area and landslide characteristic
The study area comprised a list of eight landslide events in Asia from various sources, including reports of meteorological and seismic events that triggered slope failures and field survey reports. All landslides occurred in the period ranging from 2017 to 2020, in the temporal window nominally covered by the availability of the Sentinel−1 and Sentinel−2 images. The selected study areas exhibited different environmental conditions, including slope exposition, land use, landslide size and type, and triggering factors, with the aim to improve landslide detection. The types of landslides considered were high-mobility landslides (deep-seated landslide (DSL), shallow landslide (SL), and debris flow (DF)) because they result in clear changes to the land surface after the triggering event (typhoons and earthquakes). According to the field survey and site investigation, the shallow and deep-seated landslides differed in size, geomaterial, and movement mechanism (Dai et al. 2011;Zêzere, Trigo, and Trigo 2005). Japan's Institute for Earth Science and Disaster Prevention (NIED) proposed two types of landslides: landslides with a depth greater than 10 m (deep-seated landslides) and landslides with a depth of less than 10 m (shallow landslides). Shallow landslides with surface soil mantle movement are smaller in area and volume than deep-seated landslides, which involve the movement of the surface mantle and underlying weathered bedrock (Dou et al. 2015). The selected landslide events are presented in Figures 1 and 2, and Table 1. The historical landslide data were collected mainly from scientific papers published by the Geospatial Information Authority of Japan, previous researchers, and partly from field survey reports.

Data sets and tools
This study proposes using the multitemporal amplitude ratio changes of SAR images from Sentinel−1 and differential spectral indices of optical images from Sentinel−2 with CART, which are freely available on Google Earth Engine (GEE). This approach provides a costeffective and efficient way to detect and monitor landslides over time. The SAR image and optical image datasets are presented in Table 2.

Google earth engine
Several researchers have used GEE to improve landslide detection. GEE is a cloud-based platform for satellite images, a collection of optical images (taken by Landsat, MODIS, and Sentinel−2), synthetic aperture radar (SAR) data from Sentinel−1, land cover classifications, and precipitation data, which are freely accessible datasets. GEE can achieve the integration and analysis of remote-sensing and geospatial datasets within a short period of time, using Google's cloud infrastructure (Gorelick et al. 2017). For example, high-resolution mapping of global land use and land cover classification products based on GEE have been produced in recent years (Pekel et al. 2016). Given the enormous computational load required in using  these products, they would not be feasible without GEE. In addition, pixel-based processing is facilitated by GEE, solving the problem of cloud removal in complex weather conditions, thereby substantially improving the efficiency of remote-sensing images. Moreover, GEE provides several machine-learning tools such as the random forest, classification and regression tree, support vector machine, and Naïve Bayes, which are widely used for land use and land cover classification.

Synthetic-aperture radar imagery
This study used SAR amplitude data from the Copernicus Sentinel−1 satellite constellation. The Sentinel−1 constellation launched in March 2014 and satellite has a minimum return visit time of 12 days for a specific area. Therefore, using data from satellite gives a repeat visit time of at least 12 days. Sentinel−1 carries a C-band radar sensor with a wavelength of 5.6 cm. The spatial resolution of the S1 Ground Range Detected (GRD) Interferometric Wide Swath (IW) product is 20 × 22 m. The GRD products on GEE are processed to remove thermal noise and are radiometric-and terrain-calibrated using the Shuttle Radar Topography Mission digital elevation (SRTM DEM). The image has a pixel resolution of 10, 25, or 40 m, and four polarization modes (transmission/reception sensor) are available: vertical transmission and reception (VV), horizontal transmission and reception (HH), vertical transmission and horizontal reception (VV + VH), and horizontal transmission and vertical reception (HH + HV). All Sentinel−1 GRD amplitude values provided in logarithmic decibels (dB) are calculated as 10 × log10 (A), where A is the SAR amplitude. This study used SAR data in VV and VH polarization because crosspolarizations of VV and VH are sensitive to forest biomass (Handwerger et al. 2020;Toan et al. 1992). Therefore, VV and VH can help identify landslides in vegetated areas. Furthermore, this study used multi-temporal SAR images to reduce the signal noise in the SAR data. Each stack was calculated as the temporal median of the pre-event and post-event SAR data. Preevent stacks were computed from 2015 until the time of the landslide event, whereas postevent stacks were computed within a period of 2-3 weeks after the landslide event. The SAR image stacks were constructed using a combination of ascending and descending data that were calculated as the ascending and descending means.

Optical satellite imagery
In this study, Sentinel−2 images were used to analyse pre-and post-landslide events by their spectral indices. The data were selected based on the cloud cover estimates (<10%) provided by the European Space Agency (ESA) on GEE. Sentinel−2 data provides publicly available optical satellite imagery data with the smallest ground sample distance (GSD) of about 10-60 m resolution. Sentinel−2 has multi-spectral images comprising 13 bands. The bands consist of visible, near-infrared (NIR), and short-wave infrared (SWIR) wavelengths of the electromagnetic spectrum. Additional QA60 bands are included to support the detection and removal of clouds. In this study, Sentinel−2 Multispectral Instrument Level−1C top-of-atmosphere (TOA) data were used, which exhibited radiometric and geometric corrections including orthorectification and georeferencing on a global reference system with sub-pixel accuracy. Furthermore, this study used multitemporal optical images to reduce seasonal and atmospheric noise for obtaining a pixel composite with the maximum band response, and each stack was calculated as the median of the pre-event and post-event optical images. Pre-event stacks were computed six months prior to the landslide event, while post-event stacks were computed within a period of 2-3 weeks after the landslide event.

Digital Elevation Model (DEM), topographic slope and curvature
The Shuttle Radar Topography Mission (SRTM) is an international research effort that has obtained digital elevation models on a near-global scale (Farr et al. 2007). The SRTM V3 product is provided by NASA JPL at a resolution of 1 arc-second, which is approximately a 30 m resolution. The SRTM DEM is available on GEE and can be used to calculate the topographic slope and curvature for areas unlikely to correspond to landslides (e.g. oceans, lakes, flat surfaces, and hilltops). Moreover, SRTM DEM generates topographic slope properties and curvature on GEE to remove false-positive areas. In this study, slope and curvature filtering were performed for the classification of landslide bodies, landslide deposition zones, roads, and urban areas before the landslide detection process. Breiman (1984) introduced the classification and regression tree (CART) as an effective decision tree-based approach, which has proven to be a powerful technique for handling classification or regression predictive modelling (Breiman 1984). The CART is constructed by slitting subsets of the dataset, using all predictor variables to create two sub-nodes repeatedly, and the final goal is to produce subsets of the dataset, which are as homogenous as possible with respect to the target variable. CART uses the Gini impurity index to decide which input features will provide the best split at each node (Marther and Tso 2009). The main advantage of CART is the robustness and simplicity this approach can provide for the probability of misclassification at every leaf node, thus helping to evaluate the quality of the assessment (Pal and Mather 2003). Therefore, CART can construct complex trees to solve complicated problems with large datasets (Felicísimo et al. 2013). In this study, classification using CART was performed on GEE, using the Classifer.cart approach available in the Earth Engine library.

Methodology
This study identified landslides from historical events and classified the types of landslides, using multitemporal SAR and optical images, which combined the minimum threshold of differential spectral indices and the detection of changes in the amplitude ratio with CART ( Figure 3). The methodology was developed using GEE and multitemporal satellite images (Sentinel−1 and Sentinel−2) provided by GEE. The periods of pre-and post-landslide images varied depending on the landslide event (Table 2).

Differential spectral indices by optical images
The Sentinel−2 image data of pre-and post-landslide events were filtered with a cloud mask (<10%), slope (0-20 degrees), and curvature (0.005 m −1 ) to remove the noise from clouds, shadows, landslide deposition areas, and flat areas. Landslide identification was carried out based on the loss of vegetation cover and change in land cover, using the differential normalized spectral indices. These indices have been applied to identify landslides in the target area with minimum thresholds. To detect a landslide by its spectral indices, the normalized difference vegetation index (NDVI), bare soil index (BSI), and brightness index (BI) are calculated using Equations (1) to (3).
where BLUE (band 2) is the visible blue response, GREEN (band 3) is the visible green response, RED (band 4) is the visible red response, SWIR (band 11) is the short-wave infrared response, and NIR (band 8) is the near-infrared response.
The NDVI index describes the greenness of the surface by the reflection of light waves on the ground surface. Red visible light is strongly absorbed by photosynthetically active leaves, which strongly reflect the near-infrared (NIR). A high NDVI indicates a dense forest, whereas a low NDVI represents deforested or bare soil. Moreover, the red and short-wave infrared spectral bands were used to determine the soil mineral composition, whereas the blue and near-infrared spectral bands were used to determine the presence of vegetation. The BSI captures traces of soil movements, with a high BSI representing bare soil and a low BSI representing a forested area. BI is a valuable indicator for determining the state of degradation of the soil because a low BI implies the presence of vegetation cover, whereas a high BI indicates bare soil (Ouma, Lottering, and Tateishi 2022). The BI is used to assess whether an observed surface is bright or dark, and is frequently applied for mapping soil characteristics such as roughness, texture, salinity, and moisture (Bousbih et al. 2019;Forkuor et al. 2017). However, the NDVI, BSI, and BI indices alone cannot characterize the landslide configurations because the spectral response changes quickly, especially over zones with a fast renewal of the vegetation cover after a landslide event. Therefore, the differential and relative indices were calculated to detect the landslide, using pre-landslide and post-landslide images according to Equations (4) to (6).
The results of the processing routine indicate a normalized percentage of the NDVI gained or lost. The relative difference in the normalized difference vegetation index (rdNDVI) is lost as a potential landslide and rdNDVI is gained as a non-potential landslide. The differential bare soil index (dBSI) and brightness index (dBI) take the spectral response before and after the landslide event. The area with a presence of landslides appears with negative values (Ariza et al. 2021;Scheip and Wegmann 2021).
Statistics of the spectral indices (rdNDVI, dBSI, and dBI) of the area of interest are performed to extract the value of the interval. The value of the landslide threshold is determined by means of the spectral indices and standard deviation: where Thresholds max and Thresholds min are the interval of detection. M spectrals is the mean of the spectral indices, SD spectrals is the standard deviation. To differentiate between the vegetated and bare soil pixels after landslide events, the rdNDVI, dBSI, and dBI threshold values were evaluated by pixel-based selection from eight study areas.

Detection of changes in the amplitude ratio by SAR imaging
The Sentinel−1 SAR image data examines changes in the amplitude value as a pre-landslide event stack subtracted by the post-event stack. Detection of changes in the amplitude ratio may identify the potential for landslides and other land surface changes. Furthermore, this study built the stack of pre-landslide events using Sentinel−1 images by applying a combination of ascending and descending data. This approach may reduce transient noise and errors from the atmospheric delay and other sources (Handwerger et al. 2020). Potential landslides are predicted by examining the change in amplitude, A ratio , which is defined as the pre-event stack subtracted by the post-event stack, A pre -A post ., A pre -A post is equivalent to the standard amplitude ratio approach according to Equation (9): where A pre is the amplitude of the pre-event and A post is the amplitude of the post-event. The A ratio can be either positive or negative, with positive values corresponding to a decrease in SAR amplitude after a landslide event. The SAR amplitude changes following landslide events because landslides cause substantial changes and damage to the ground surface, which alter the radar reflectance, hillslope geometry, roughness, and dielectric properties of the ground before and after the landslide. The SAR amplitude stacks of pre-events images could improve landslide detection that were found by (Adriano et al. 2020;Handwerger et al. 2020).

Combining optical-and SAR-based classification by CART
The results of landslide detection using differential spectral indices with the minimum threshold (Sentinel−2) and detection of amplitude ratio changes (Sentinel−1) were applied to CART for evaluating the identification of landslides during the classification phase. The combination of optical and SAR images may overcome their limitations, including atmospheric and meteorological conditions, speckle noise, low coherence in dense vegetation, and others. The differential spectral indices and amplitude ratio changes from historical landslides in eight study areas were randomly divided into training data (70% of landslide locations: 171 landslides) and validation data (30% of landslide locations: 74 landslides). The spatial resolution for landslide detection was 10 m (as input pixel spacing of 10 × 10 m). The accuracy assessment used the F1-score to evaluate the performance of the model. This method combined precision and recall to assess the performance of the landslide detection model.

Accuracy assessment
The performance of the landslide detection approach was evaluated by applying known classification metrics, which were also used to evaluate the accuracy of the model. These classification metrics include accuracy, precision, recall, and F1-score. Because overall accuracy is very sensitive to the data distribution (He and Garcia 2009), the detection of numerous true negatives considerably increases the accuracy of the classification. Therefore, the impact of the minority class (target class) is reduced compared to that of the majority class (Branco, Torgo, and Ribeiro 2017). Therefore, the accuracy of the model should be evaluated to eliminate the sensitivity of true negative detection and estimate the precision, recall, and F1-score. The following equations show how these metrics are calculated: where TP is a true positive (a landslide is correctly identified), TN is a true negative (a nonlandslide is correctly identified), FP is a false positive (a non-landslide is incorrectly identified as a landslide), and FN is a false negative (a landslide is incorrectly identified as a non-landslide). The precision and recall evaluate different aspects of the method; thus, an index that combines both is also used. The F1-score is the harmonic mean of the precision and recall; its highest value is 1 (100%) and its lowest is 0 (0%).

Classification of landslide types with a threshold value
In this study, the types of landslides (shallow and deep-seated landslides) were classified using a combination of optical-and SAR-based classification by CART (section 4.1.3). The analysis was based on the relationship between the differential spectral indices and amplitude ratio changes. In addition, the classification was based on changes to the ground surface, roughness, and source material in postlandslide areas because these factors substantially affect spectral reflectance, radar reflectance, and scattering. The study focused on three landslide events in Ehime, Hokkaido, and Kumamoto, where shallow and deep-seated landslides were observed during the same event. The landslide trigger was rainfall in the Ehime and Kumamoto events, whereas a combination of rainfall and an earthquake triggered the Hokkaido event. Finally, the accuracy of the landslide classification was evaluated using the F1-score.

Landslide detection model
The proposed framework analysed the landslide detection results in eight study areas available on GEE (Figure 3). For the landslide detection process, the multitemporal SAR and optical images were calculated as the pre-and post-event temporal median for differential spectral and amplitude ratio change analysis. The thresholds of the differential spectral indices are essential for identification, and the statistical analysis of rdNDVI, dBSI, and dBI was applied to extract the interval value, using Equations (7) and (8). The differential spectral index data for threshold analysis were obtained from each pixel value on different types of land cover (landslide, deposition area, road, urban area, and forest) in eight study areas (Figure 4). The minimum thresholds of rdNDVI, dBSI, and dBI were set to −100, −15, and −0.02, respectively, based on the visual evaluation from the eight study areas. Furthermore, the relationship between rdNDVI, dBSI, dBI, and the change in amplitude ratio with a slope was defined according to the type of land cover. The forested or vegetated areas were defined as having positive rdNDVI, dBSI, and dBI values because the land cover did not change or changed only slightly during the landslide events. By contrast, the landslide and deposition areas were defined as having negative values as a result of the substantial changes in land cover and surface properties that altered spectral reflectance. The amplitude ratio change of the landslide was greater than that of the forested or vegetated areas because the landslide caused notable modifications to the surface and deforestation, which decreased the backscattering reflectance to the satellite compared to the prelandslide surface (Adriano et al. 2020). Hence, the differential spectral indices and amplitude ratio changes differed according to the landslides, deposition areas, and forested areas. Moreover, the landslide and deposition areas were classified by slope filtering. The slope of the landslides was above 10 and the slope of the deposition areas was below 10.
The minimum thresholds of the differential spectral indices of the eight study areas were applied to CART on GEE for landslide detection. Figure 5 and 6 show the results of the proposed landslide detection model. The accuracy of the assessment was obtained from the total number of landslides and areas analysed by CART, based on a semi-automatic classification. These results considered different parameter classifications according to two conditions, the first of which was analysed by applying all differential spectral indices (rdNDVI, dBSI, and dBI), and the second of which was analysed by applying all differential spectral indices and the detectioFn of amplitude ratio changes (rdNDVI, dBSI, dBI, and A ratio ). The accuracy of the proposed model for landslide detection was rated as moderately good to excellent ( Table 3). The accuracy of the model obtained an F1-score of 70.01-94.55% for the landslide detection using all differential spectral indices by Sentinel−2 ( Figure 5) and 79.60-96.38% for the landslide detection using all differential spectral indices and the detection of the amplitude ratio changes ( Figure 6). Furthermore, the F1scores of the landslide detection models using rdNDVI, dBSI, dBI, and A ratio were 41.27-77.65% for rdNDVI, 42.44-92.32% for dBSI, 46.53-78.01 for dBI, and 40.04-  72.15% for A ratio . Regarding the accuracy of the results, the detection of amplitude ratio changes improved the landslide detection model as revealed by the F1-score. In addition, a quantitative comparison of landslide detection using several indices is illustrated in Figure 7. The outcome parameters indicated that the landslide detection model utilizing rdNDVI, dBSI, dBI, and amplitude ratio changes showed low TP and high FP rates when applied individually because specific indices could not be applied to all study areas due to environmental conditions. However, the combination of differential spectral indices increased TP and decreased FP because of the spectral index compensation, as demonstrated by the landslide detection model that used all differential spectral indices. Furthermore, integrating differential spectral indices with amplitude ratio changes substantially enhanced TP and reduced     FP. The detection of amplitude ratio changes eliminated FP in the vegetated and urban areas, which did not show substantial amplitude changes in the pre-and post-landslide SAR images. For the Hiroshima, Ehime, Yunnan, and Kumamoto landslide events, all differential spectral index models detected false pixels in urban areas and roads. Nevertheless, the model combining all differential spectral indices with amplitude ratio changes detected fewer false pixels in the urban areas and roads. For the Hokkaido, Kyushu, Chiang Rai, and Brebes landslide events, the model using all differential spectral indices combined with amplitude ratio changes eliminated false pixels in vegetated areas. Moreover, the multi-temporal optical images stack can reduce seasonal and meteorological noise and the detection of changes in the amplitude ratio was analysed with a stack of pre-event images to reduce noise from radar backscattering, spots, geometry shade, and shadow in the SAR data. Hence, this type of detection (Sentinel−1) overcame the limitations of optical images (Sentinel−2), which required sunlight and cloud/shadow-free conditions to identify landslides. However, the detection of changes in the amplitude ratio using Sentinel−1 images cannot be applied to landslide detection by itself because C-band SAR images are sensitive in vegetated areas and large changes in the water content due to rainfall influence SAR reflectivity (Adriano et al. 2020;Jung and Yun 2020). Such errors were observed mainly in forested areas and grasslands but could be mitigated by using a longer wave radar, which has a greater ability to penetrate vegetated areas (Schlögel et al. 2015). The accuracy of landslide detection by the amplitude ratio changes obtained a score of 40.04-72.15%. The accuracy of the model using all differential spectral indices together with the amplitude ratio changes showed a score 3.60% higher than that of the model that omitted the amplitude ratio changes. Therefore, the combination of spectral indices and amplitude ratio changes for landslide detection reduced noise and false pixels. Therefore, this combined model could be useful for improving landslide detection and could be applied to other Asian regions. Moreover, slope filtering reduced FP by removing some areas (roads, urban areas, lakes, oceans, flat surfaces, and hilltops) as shown in Figure 8. The topographic slope and curvature were calculated by 1 arc-second resolution SRTM DEM. The slope filtering classified landslides and the surrounding environment ( Figure 3). The slope filtering from 2.5 to 20 degrees examined the landslide detection model, which increased accuracy by more than 10 degrees (Figure 8). This outcome is related to the slope characteristics of landslides, in which the landslide results in a steep slope and a hill slope, as shown in Figure 4. Therefore, the combination of differential spectral indices and amplitude ratio changes using multi-temporal images stack improved the landslide detection model, and the slope filtering enhanced the accuracy of the model.

Landslide types classification
The classification of landslides (deep-seated and shallow landslides) was analysed by considering the relationship between the differential spectral indices and amplitude ratio changes. Figure 9 shows the relationship between the differential spectral indices and amplitude ratio changes in the Ehime and Kumamoto events, which were two types of landslides triggered by rainfall. The study determined that the minimum thresholds for rdNDVI, dBSI, and dBI were less than −30, −300, and −0.05, respectively, while the amplitude ratio change was greater than 1.0 for identifying deep-seated landslide. The differential spectral indices and amplitude ratio changes of the deep-seated landslide were greater than those of the shallow landslide because the deep-seated landslide caused considerable damage on the ground surface, exhibited different soil texture and mineral composition, and changed the depth of the ground surface of the post-landslide body (Figure 10 and 11). Therefore, the deep-seated landslide and shallow landslide were classified by    applying the rdNDVI, dBI, dBSI, and amplitude ratio changes. These indices detected the different soil textures and weathered rock material as well as the depth of the post-landslide body as these changes corresponded to the reflectance of different spectral bands. Figures 10 and 11 show the magnitude of different spectral indices and amplitude ratio changes in the Ehime and Kumamoto events, which showed similar trends as deep-seated landslides and substantially differed from shallow landslides. However, the relationship between the differential spectral indices and amplitude ratio changes of deep-seated and shallow landslides in Hokkaido differed from those in Ehime and Kumamoto. The distribution of differential spectral indices and changes in the amplitude ratio was the same for both deep-seated and shallow landslides because an earthquake was the trigger ( Figure  12 and 13). The earthquake caused rapid and substantial changes to the surface of an extended area as well as considerable damage to the vegetated area, which may have affected the radar reflectance and backscattering distribution.
The minimum thresholds of the differential spectral indices and amplitude ratio changes ( Figure 9) were applied to the landslide detection framework for classifying the landslides. Figures 14 and 15 show the detection results of the deep-seated landslides in Ehime and Kumamoto, using the classification. Deep-seated landslides were observed at both events but shallow landslides were not observed because the differential spectral indices of shallow landslides were lower than those of deep-seated landslides. For Hokkaido events, the deep-seated and shallow landslides could not be classified because the differential spectral indices were similar ( Figure 16). Moreover, the F1 accuracy scores were 88.41% and 90.55% for the detection of the deep-seated landslides in Ehime and Kumamoto, respectively. On the other hand, the F1-score of the Hokkaido event was 33.91%, which suggests poor performance by the deep-seated landslide detection model in the case of an earthquake trigger (Table 4). Hence, the threshold of the differential spectral indices and amplitude ratio changes was able to classify deep-seated landslides and shallow landslides triggered by rainfall, and the trigger could affect the  classification of landslides depending on the differential spectral indices and detection of amplitude ratio changes as shown by the Ehime, Kumamoto, and Hokkaido events.

Conclusion
The landslide detection model proposed in this study was developed to identify landslides using Sentinel−1 SAR data and Sentinel−2 optical data available on the Google Earth Engine (GEE), which are freely available data that can be used to map the landslides after catastrophic events. The accuracy of the model was validated using the F1-score equation with historical landslide data from eight study areas in Asia. The results showed that the landslide detection model successfully detected landslides with F1-scores of 70.01-94.55% for the model using all differential spectral indices by Sentinel−2, and 79.60-96.38% for the model using all differential spectral indices and amplitude ratio change detection. The accuracy of the model for landslide detection was considered moderately good to excellent. Moreover, the amplitude ratio changes provided by Sentinel−1 SAR could be helpful for removing false pixels in vegetated and urban areas. Slope-filtering during pre-image processing classified landslides and the surrounding environments according to their different slope characteristics by removing false positives. In addition, the minimum thresholds of differential spectral indices and amplitude ratio change were applied to detect landslides for eight study areas and the detection result were moderately good and excellent.  Landslides were classified into two types (deep-seated and shallow landslides) using the thresholds of the differential spectral indices (rdNDVI, dBSI, and dBI). The differential spectral indices of the deep-seated landslide were greater than those of the shallow landslide because of the greater changes to the ground surface and texture of the material after the landslide event. These indices detected the different textures of the soil and weathered rock material in the post-landslide body, which affected radar reflectance and change detection. Classification of landslides using the deep-seated landslide threshold was successful for the Ehime and Kumamoto events, which were triggered by rainfall. However, the model could not perform a classification for the Hokkaido events, which were triggered by rainfall and an earthquake. The earthquake caused rapid and extensive changes to the ground surface as well as substantial damage to the vegetated area, which may have affected the radar reflectance and backscattering distribution.
Although the aims of this study were achieved, this study showed limitations as a result of the rapid response during events because the landslide detection model in this study required a cloud cover of less than 10% for the Sentinel−2 images. Generally, clouds cover the landslide area during rainfall, and good-quality images are obtained 2 to 4 weeks after the landslide event. Another limitation was the SAR data from Sentinel−1 because C-band satellite images are sensitive to vegetated areas. Unfortunately, ALOS−2 is an L-band imaging satellite that can be effective in vegetated areas but its images are rarely used worldwide and are not freely available. In the future, the NASA-ISRO SAR (NISAR) mission will operate with the L-band SAR sensor, which generally produces better images for vegetated areas, and will be available to the public as the Sentinel program.