Enhancing landslide detection in Western Ghats of Kerala, India with deep learning and Explainable AI Authors: Abhilash Sreekumar, Hemalatha Thirugnanam, Sansar Raj Meena, Maneesha Vinodini Ramesh Published: 2025 ABSTRACT The Western Ghats in Kerala, India, have witnessed a significant rise in landslide occurrences in recent years. More recently in 2024, catastrophic landslides in Wayanad district claimed 336 lives, with 78 people reported missing, underscoring the urgent need for accurate prediction and effective risk mitigation strategies. A reliable historical landslide database is critical to address this escalating challenge. While deep learning has been successfully employed in various studies for creating landslide inventory, with some of them being in Western Ghats itself, an automated landslide inventory is still not available for the state of Kerala in India. In this study, temporal differences from pre and post-event high-resolution satellite imagery is used to train a deep learning model for landslide segmentation. PlanetScope imagery with a 3-meter resolution, along with a publicly available landslide database made from it named High Resolution Landslide Detector Database (HR-GLDD) were utilized for this. Two variants of the U-Net model – simple multiscale U-Net and attention multiscale U-Net were trained on HR-GLDD and fine-tuned with temporal difference data using transfer learning. The efficiency of both models in landslides mapping was evaluated quantitatively using metrics and qualitatively using visual comparison of outputs. Explainable Artificial Intelligence (XAI) techniques such as Integrated Gradients (IG) and Gradient-Weighted Class Activation Mapping (Grad-CAM), as well as feature map visualization are used to understand further about the model’s functioning. The attention multiscale U-Net model achieved best performance with a precision of 91.49% and an F1 Score of 81.69%. Visual comparison of predictions with ground truth further demonstrated the effectiveness of the attention multiscale U-Net in accurate inventory mapping. IG identified the Near-Infrared (NIR) band as the most influential input for landslide segmentation, while Grad-CAM and feature map visualization helped to reveal spatial patterns captured by the model at different stages. This study can be seen as a successful first step towards building an automated landslide inventory for the state of Kerala in India. Introduction Landslides, driven by geological, hydrological, and anthropogenic factors, pose a growing threat in Kerala, India, necessitating accurate detection and mitigation strategies1,2. Accurate historical landslide databases play a pivotal role in enhancing landslide susceptibility modeling and land use decision-making processes3. Conventional methods for generating landslide inventories often depend on collating inventories from historic records, field investigations and aerial photo interpretation which are time-consuming, costly, and may not be feasible for large or inaccessible areas1. Advancements in remote sensing technologies, coupled with the increasing availability of high-resolution satellite imagery, and deep learning techniques have emerged as powerful tools for automated landslide detection and mapping, demonstrating superior performance compared to traditional machine learning algorithms due to their ability to automatically learn complex patterns and features from large datasets1,4,5. Numerous studies have explored landslide monitoring techniques in the past. A 2022 review examined these techniques adopted in different parts of the world, highlighting the challenges and opportunities in integrating Internet of Things (IoT) into these methods6. A wide array of machine learning and deep learning models have been used in the past for mapping landslides. These models and architectures were reviewed in 2023 along with their associated challenges, constraints, and recommendations to overcome them7. Several studies have put efforts towards mapping susceptibility of landslides in landslide prone regions. A 2023 study investigated the impact of land-use land-cover change in landslide susceptibility of Idukki district in Kerala8. Another study done in the same year investigated the relationships between properties of lithological units (such as cohesion, friction, angle, and rock formation) and their role in earthquake-triggered landslides in specific regions of India, Taiwan, China, and Japan9. An evaluation of the vulnerability of landslides & rockfalls in Tirumala hills of Andhra Pradesh, India using a catchment based approach was done in 202310. Another study done in the same year used the Markov-Switching Spatio-Temporal Generalized Additive Model for landslide susceptibility mapping11. Traditional landslide inventories, constrained by missing data, inaccessible terrains, and resource-intensive processes, limit comprehensive mapping1,3. High-resolution satellite imagery such as PlanetScope and Sentinel-2 improves mapping efficiency, yet spectral similarity challenges persist4,5. Other datasets used in landslide studies include Landsat12,13 LISS IV imagery14, Pleiades imagery13,15, GeoEye1 imagery16, SPOT imagery, and Worldview 2 imagery15. The timeline of important multispectral satellite missions, along with their spatial and temporal resolution characteristics, is illustrated in Fig. 1. Fig. 1Timeline of key multispectral satellite missions. The operational period of each mission is represented by the length of their corresponding bar. Bar colour indicates spatial resolution, and style of the bar borders illustrate the revisit time. Timeline of key multispectral satellite missions. The operational period of each mission is represented by the length of their corresponding bar. Bar colour indicates spatial resolution, and style of the bar borders illustrate the revisit time. Numerous efforts have produced publicly available landslide inventories. The spatial distribution and publication timeline of these datasets are shown in Fig. 2. Fig. 2Global distribution and publication timeline of key landslide inventory datasets. (a) Worldmap showing the geographic distribution of publicly available landslide inventories. Each point’scolor corresponds to a specific dataset, while point size reflects the number of landslidesrecorded. (b) Timeline displaying the publication year of each dataset. Global distribution and publication timeline of key landslide inventory datasets. (a) Worldmap showing the geographic distribution of publicly available landslide inventories. Each point’scolor corresponds to a specific dataset, while point size reflects the number of landslidesrecorded. (b) Timeline displaying the publication year of each dataset. A 2020 study developed an inventory of landslides from the 2018 Kerala monsoon event14. Landslides triggered by the 2008 Wenchuan Earthquake have also been utilized in research15. Additionally, several studies have used publicly available landslide datasets from Nepal including Bijie landslide inventory12,17. A 2022 study created landslide inventory for Patagonian Andes, addressing the absence of a dedicated dataset for region18. In 2022, researchers developed the Landslide4Sense dataset, which serves as a benchmark dataset for landslide detection, incorporating multispectral and topographical data from over 4800 landslides across multiple regions19. Landslide Recognition along the Sichuan-Tibet Transportation Corridor (LRSTTC) dataset, an optical remote sensing dataset designed for landslide detection along the Sichuan-Tibet Transportation Corridor is also being utilized by landslide studies20. A 2023 study introduced the High-Resolution Global Landslide Detector Database (HR-GLDD), which integrates PlanetScope tiles of landslides from 10 different physiographic regions worldwide21. Additionally, a 2024 study mapped landslides from the Qinghai-Tibet Plateau region in China and made the data publicly available20. Another study in the same year compiled an inventory for the Himalayas, documenting recurrent and persistent landslides spanning a 30-year period22. The emergence of deep learning algorithms has significantly enhanced automated and semi-automated approaches for landslide mapping in the geospatial domain. Numerous studies have employed such algorithms for creation of landslide inventories as well as for susceptibility mapping23–25. Among these, Convolutional Neural Networks (CNNs), including variants like LanDCNN and Mask Region-based CNN (Mask R-CNN), have proven effective for landslide mapping13,20,26–28. Among the computer vision algorithms similar to CNN, U-Net is one of the most widely used and proven to be effective models in image segmentation tasks, since its deconvolution and skip connection improves localization and minimizes loss of information. Several studies using the U-Net model for landslide detection and susceptibility mapping have been published since 20193,12,24,25,29,30. Several studies have explored Res U-Net, a modified version of the ordinary U-Net al.so for landslide mapping16,31. Other variants and modifications include Siamese Nested U-Net (SNUNet)32, U-Net + +33, Attention Deep Supervision Multi-Scale U-Net (ADSMS U-Net)34, U-Net 3+, and Transformer U-Net (TransU-Net)15, as well as U-Net with a Residual Neural Network (ResNet) backbone22. A 2022 study employed 11 different deep learning models for landslide mapping and compared their performances, in which Residual U-Net (ResU-Net) achieved superior performance19. A 2024 study used visual saliency data to improve landslide mapping outcomes35. Landslide inventories created using such methods are crucial for generating the landslide susceptibility map of the regions. However, some of the major issues in susceptibility mapping models is their poor generalization, high uncertainty in model performance, and inaccurate selection of negative samples (non-landslide samples). To overcome these issues, physics-informed machine learning approaches have been applied, and have transformed landslide mapping studies. One such model is Transient Rainfall Infiltration and Grid-based Regional Slope-stability Analysis (TRIGRS), which was integrated with CNN in a 2021 study for reducing prediction uncertainty and improving generalizability of the model in susceptibility mapping. The proposed model saw a 4.9% increase in Area Under Curve (AUC), and 24.8% drop in uncertainty36. A 2023 study compared various hybrid data-driven and physical models in terms of generalization and prediction uncertainty across different regions, in which CNN integrated with Infinite-Slope Stability Model (ISSM) gave optimum performance with an AUC of 83.4%37. A study published in 2024 combined Random Forest (RF) model with Scoops 3D model for optimum negative sample selection in landslide susceptibility mapping through slope stability analysis, achieving 36.7% increase in AUC38. The combination of RF and Scoops 3D was also used by another study in the same year, where CNN was employed for modelling bedrock depth of the study area, which was then used as a conditioning factor for landslide susceptibility mapping39. Another 2024 study used CNN with Physically-based Probabilistic Model (PPM) for reducing uncertainties in susceptibility mapping of rainfall-induced landslides, where the proposed model achieved 85% AUC40. Despite these advancements and innovations, there are still challenges in the domain of landslide mapping and inventory creation. Misclassification of spectrally similar features such as built-up areas, barren land and riverbeds is one of the major challenges while creating landslide inventories using machine learning and deep learning algorithms. Several studies have tried including pre-event and post-event images to train the model in order to overcome this26,32,41. To improve mapping accuracy and model performance, studies often incorporate auxiliary datasets and derived variables. Topographical data such as elevation and slope derived from Digital Elevation Models (DEM), Digital Terrain Models (DTM), and Digital Surface Models (DSM), are widely used to identify terrain characteristics that influence landslide occurrence19,22,25,27,28,30,42. Additionally, the Normalized Difference Vegetation Index (NDVI) derived from the multispectral imagery has been utilized to identify vegetation disturbances that may indicate landslide activity, especially in forest regions3,42. A 2025 study incorporated night time light data into the training dataset for landslide mapping22. Limited data for training and testing remains another significant challenge faced when training deep learning models for landslide mapping. Several studies have adopted transfer learning to address this issue, where an already available large dataset can be leveraged to train a model while a small dataset from the study area can be used to fine-tune the model weights18,20,34. Landslides obscured by dense vegetation and other local factors are also difficult to map automatically34,43. Using microwave remote sensing datasets such as Synthetic Aperture Radar (SAR), which can penetrate through vegetation canopy, can help with this issue to an extent. Another disadvantage is that deep learning models tend to merge multiple landslides in close proximity as a single landslide3,27. These are challenges that are still being researched and need to be solved by future studies. Although bitemporal imagery and their temporal differences were used in regions such as China26,28 and Himalayas32 for improving model predictions, it has been seldom used for landslide inventory creation in Kerala. Explainable Artificial Intelligence (XAI) techniques such as Shapley Additive Explanations (SHAP) have been used in several landslide studies for susceptibility mapping with transparency in machine learning and deep learning algorithms44,45, but it was not previously implemented for inventory mapping especially in Western Ghats region in Kerala. Therefore, as a first step towards building an automated landslide inventory for the region, this study aims to deploy deep learning algorithms for mapping landslides from Western Ghats of Kerala using temporal differences from high-resolution bitemporal satellite imagery. The study also uses XAI techniques and feature map visualization as supportive tools to interpret model behavior and to understand the features learned by it. We are focusing on adaptation of existing methods for a new, data scarce region rather than developing new methods or improving upon the existing methods. Study area The study area selected for this research is a region in the Western Ghats of Kerala, India, located north of the Palakkad Gap. It spans across the districts of Kozhikode, Wayanad, Malappuram, and Palakkad as shown in Fig. 3. The total area of the selected region is 2,938.68 km2, with a latitudinal extent ranging from 10.8°N to 11.95°N and a longitude extent from 75.68°E to 76.84°E. Maximum elevation within this area reaches 2,553 m as derived from the Shuttle Radar Topographic Mission (SRTM) DEM. Selection of this area was based on its slope characteristics, geomorphology, and frequency of historic landslides in it. Specifically, this area comprises a spatially contiguous region with slopes equal to or greater than 20° across the districts of Kozhikode, Wayanad, Malappuram, and Palakkad, and is covered by moderately and highly dissected hills, valleys and plateaus as shown in Fig. 4. It is also a hotspot of landslides in Kerala, with around 25.9% of recorded landslides in the state being here as per the inventories from Bhukosh, Geological Survey of India (GSI)46 and van Westen (2020)47. Furthermore, the devastating Mundakkai landslide of 2024 that saw a staggering 401 deaths also happened in this region48. Some of the landslides that happened in the area along with their brief details are given in Table 1. Fig. 3Location map of the study area - the Western Ghats region to the north of Palakkad gap across the districts of Wayanad, Kozhiikode, Malappuram, and Palakkad. Basemap is taken from Google Terrain in Quickmap services of QGIS. Elevation data is taken from SRTM DEM. Location map of the study area - the Western Ghats region to the north of Palakkad gap across the districts of Wayanad, Kozhiikode, Malappuram, and Palakkad. Basemap is taken from Google Terrain in Quickmap services of QGIS. Elevation data is taken from SRTM DEM. More than 52.9% of the region is covered by clay soil, while gravelly clay covers more than 27.85% of the area. The remaining area consists of gravelly loam and loam which consist of 13% and 5.84% respectively49. From a geological perspective, about 48.5% of the area is covered by charnockite, while migmatite and peninsular gneissic complexes cover around 27.9% and 15% respectively50. Rest of the area, which is less than 10%, is covered by acid intrusives, Wayanad group, Satyamangalam group, intermediate intrusives, khondalite, and gabbro-anorthosite complex50. As per Koeppen’s scheme of climate classification, most of Kerala is grouped under tropical monsoon type, which is characterized by two monsoon seasons: South-West monsoon (June to September) and North-East monsoon (October to November). The long-term average annual rainfall in the state in the period 1871 to 2018 was 282 cm51. Fig. 4Geomorphology and historic landslides in the study area. Concentration of landslides is represented across latitudes and longitudes using line charts. Geomorphological data was obtained from Bhukosh, GSI52, while information on past landslide events was acquired from Bhukosh, Geological Survey of India (GSI)46 and van Westen (2020)47. Geomorphology and historic landslides in the study area. Concentration of landslides is represented across latitudes and longitudes using line charts. Geomorphological data was obtained from Bhukosh, GSI52, while information on past landslide events was acquired from Bhukosh, Geological Survey of India (GSI)46 and van Westen (2020)47. Data preparation Data collection For preparing training and testing samples, cloud-free PlanetScope imagery for the period 2018–2025 was acquired from Planet Labs Inc. This dataset provides high resolution multi-spectral images with 3 m spatial resolution in four bands: blue (455–515 nm), green (500–590 nm), red (590–670 nm), and Near-Infrared (NIR) (780–860 nm). Table 1Details of selected major landslide events happened in the study area and were captured by the model developed in this study.PlaceDateCoordinatesRemarksMundakkai, Chooralmala, Vellarimala, Meppadi (Wayanad District)30th July, 202411° 28’ 01” N, 76° 08’ 09” EAround 336 people died, with 78 people reported missingPuthumala, Kalladi, Vythiri (Wayanad District)08th August, 201911° 29’ 51” N, 76° 07’ 47” E11 people killed in the incident, 6 people missingKavalappara, Pothukal, Nilambur (Malappuram District)08th August, 201911° 24’ 32” N, 76° 14’ 11” E59 people killed in the landslidePathar, Pothukallu (Malappuram District)08th August, 201911° 23’ 25” N, 76° 13’ 32” E16 houses destroyed, 30 houses partially damagedIrumbakachola, Mannarkad (Palakkad District) 01 st September, 202111° 00’ 44” N, 76° 45’ 06” EAgricultural land was damaged.Kannappankundu, Thamarassery (Kozhikode District)08th August, 201811° 32’ 04” N, 75° 59’ 13” EOne person killed, 10 houses destroyed, 20 houses partially damaged. Table 1Details of selected major landslide events happened in the study area and were captured by the model developed in this study.PlaceDateCoordinatesRemarksMundakkai, Chooralmala, Vellarimala, Meppadi (Wayanad District)30th July, 202411° 28’ 01” N, 76° 08’ 09” EAround 336 people died, with 78 people reported missingPuthumala, Kalladi, Vythiri (Wayanad District)08th August, 201911° 29’ 51” N, 76° 07’ 47” E11 people killed in the incident, 6 people missingKavalappara, Pothukal, Nilambur (Malappuram District)08th August, 201911° 24’ 32” N, 76° 14’ 11” E59 people killed in the landslidePathar, Pothukallu (Malappuram District)08th August, 201911° 23’ 25” N, 76° 13’ 32” E16 houses destroyed, 30 houses partially damagedIrumbakachola, Mannarkad (Palakkad District) 01 st September, 202111° 00’ 44” N, 76° 45’ 06” EAgricultural land was damaged.Kannappankundu, Thamarassery (Kozhikode District)08th August, 201811° 32’ 04” N, 75° 59’ 13” EOne person killed, 10 houses destroyed, 20 houses partially damaged. Details of selected major landslide events happened in the study area and were captured by the model developed in this study. Kavalappara, Pothukal, Nilambur (Malappuram District) The imagery also has a pixel depth of 16-bit, and a near-daily revisit cycle which allows frequent data acquisition. High Resolution Global Landslide Detector Database (HR-GLDD) is a dataset composed of 1,758 landslides, manually digitized from 10 different physiographic regions worldwide. The input layer consists of post-event multispectral PlanetScope images of landslides, and the corresponding output layer is the binary mask where pixels associated with landslides are labeled as 1, and non-landslide pixels are labeled as 0. This dataset, created by Meena et al. (2022)21, is publically available for research in numpy array format (.npy) and can be accessed53. Since this study aims to develop a model for mapping landslides from PlanetScope imagery, HR-GLDD is leveraged for the initial training of the model. Data preprocessing The methodology used for data preparation and model training are given in Fig. 5. Fig. 5Flowchart illustrating methodology used for this study. Blue colored shapes indicate data preprocessing steps while green colored shapes indicate model training and evaluation steps. Parallelograms indicate input datasets or outputs; rectangles indicate processes or actions; hexagons indicate models. Numbers given in parenthesis inside parallelograms indicate the number of samples in the corresponding dataset. Flowchart illustrating methodology used for this study. Blue colored shapes indicate data preprocessing steps while green colored shapes indicate model training and evaluation steps. Parallelograms indicate input datasets or outputs; rectangles indicate processes or actions; hexagons indicate models. Numbers given in parenthesis inside parallelograms indicate the number of samples in the corresponding dataset. A total of 444 samples (222 landslide and 222 non-landslide features) were digitized from the study area with ArcGIS Pro and saved in GeoJSON format. These samples were used to extract 128 × 128 shaped tiles for each sample from pre-event and post-event PlanetScope imagery. From these, a new temporal difference layer was generated for each sample by performing band-wise subtraction between the corresponding pre-event and post-event tiles. This helps to enhance landslide detection by highlighting landslides, which are present only in post-event images, while suppressing unchanged and stable features present in both pre-event and post-event images such as barren land & riverbed. These features are often spectrally similar to landslides; therefore their suppression helps to minimize false positives during landslide segmentation. After generating the temporal difference dataset with 444 samples, it was split into two subsets: ~75% for training (332 samples) and ~ 25% for testing (112 samples). This ratio for train-test split was chosen since it gives training data large enough for the model to generalize, and testing data with enough samples to properly evaluate the model performance. This ratio is used in several previous studies as well. Both subsets underwent preprocessing steps, including normalization and data augmentation, to enhance model performance and generalization. In normalization, the pixel values in each band of the layer are scaled into the range 0 to 1 for easier computations and a stable training process. In data augmentation, the number of instances in the dataset is increased by flipping and rotating the samples in it. Specifically, six augmentation methods were applied to each sample: rotations of 90°, 180°, and 270°, as well as horizontal and vertical flipping. This process increases the dataset size from 112 to 672 in the testing set, and 332 to 1992 in the training set, with 5 augmented samples generated out of each original sample (from 444 to 2664 in total). Model implementation This study uses 2 different variants of the U-Net model for landslide segmentation. U-Net is a deep learning algorithm designed to highlight particular objects or regions in an image, making it widely applicable to image segmentation tasks. The architecture follows an encoder-decoder framework, commonly referred to as an autoencoder. This structure efficiently learns spatial hierarchies and maps pixel-level relationships within an image. Encoder progressively reduces the spatial resolution of the input image through downsampling while extracting the most significant features and learning important patterns. These features are represented in the form of feature maps, stored in the channels of the encoder. Decoder, in contrast, restores the spatial resolution of the image by upsampling the feature maps. It determines the precise spatial location of the features extracted by the encoder, enabling the reconstruction of the output image with the desired segmentation. The information extracted by each encoder block is passed on to its respective decoder block through skip connections or concatenation layers to prevent loss of information while downsampling. In the end, the features learned by the model are combined to produce the final output in the form of a binary mask that shows the features predicted by the model from input. The U-Net variations used in this study are: simple multiscale U-Net, and attention multiscale U-Net. Simple multiscale U-Net The multiscale U-Net architecture designed for this study is illustrated in Fig. 6. It has 4 encoder-decoder blocks separated by one bottleneck. In each block, the input goes through 2 convolution layers, each followed by a batch normalization and Rectified Linear Unit (ReLU) activation. The dimensions of model input is reduced from 128 × 128 in the first block to 8 × 8 in the bottleneck block while extracting features in each stage. In each encoder block, the input image is progressively downscaled through average pooling to match the resolution required at that level. These scaled inputs are then passed through convolution layers to extract scale-specific features, which are then concatenated with the feature maps collected from the previous block. This enables integration of contextual information from multiple spatial scales. The features collected at the end of each encoder block are passed on to the respective decoder through concatenation as well. There are 32 filters in the convolution layers of the first block, which gives 32 feature maps. The second, third, fourth, and bottleneck blocks have 64, 128, 256, and 512 such filters respectively. Attention multiscale U-Net In attention multiscale U-Net, attention gates are integrated with the simple multiscale U-Net to improve the model’s ability to focus on the most relevant regions in an input image. This is done by suppressing irrelevant or noisy features while highlighting the important ones, thereby improving feature selection and overall model performance. The attention gate achieves this by creating a resampler (attention map) through linear transformation, element-wise addition, non-linear activation, and normalization of feature maps from encoder decoder blocks. This resampler is subsequently applied to the encoder features via element-wise multiplication, which allows the model to focus on the most important features. The attention multiscale U-Net used for this study is illustrated in Fig. 7. It has 16, 32, 64, 128, and 256 filters in the first, second, third, fourth, and bottleneck blocks respectively, extracting the same number of feature maps in the corresponding stages while taking contextual information from the downscaled input image. Fig. 6Simple multiscale U-Net architecture. The numbers above each block represent the number of feature channels (bands), and the numbers on sides denote the spatial dimensions of the corresponding feature maps. Simple multiscale U-Net architecture. The numbers above each block represent the number of feature channels (bands), and the numbers on sides denote the spatial dimensions of the corresponding feature maps. Fig. 7Attention multiscale U-Net architecture. The numbers above each block represent the number of feature channels (bands), and the numbers on sides denote the spatial dimensions of the corresponding feature maps. Attention multiscale U-Net architecture. The numbers above each block represent the number of feature channels (bands), and the numbers on sides denote the spatial dimensions of the corresponding feature maps. Model training and evaluation The models designed for this study were initially trained using the HR-GLDD dataset. Subsequently, the initial layers of the models were frozen, and the rest of the layers were fine tuned using the temporal difference dataset created from the study area. This process, where a model pre-trained on a large dataset is fine-tuned using a similar, but different, typically smaller dataset is known as transfer learning. It allows efficient training of deep learning models on small datasets created from the study area by leveraging publicly available, larger, similar datasets. However, a key limitation of the transfer learning approach is that the number of layers (bands) must remain consistent between the pre-training dataset and the dataset used during the transfer learning phase. Since the HR-GLDD only contains 4 spectral bands from post-event imagery, we could not incorporate additional auxiliary variables such as vegetation indices or terrain derivatives. To maintain compatibility with the pre-trained weights, the study area dataset was restricted to the same number of bands, but with temporal differences between pre-event and post-event images instead of just post-event layers. After model training, evaluation was done quantitatively using precision, recall F1 Score, IoU Score, and MCC. Qualitative evaluation was done via visual comparison of the model outputs with ground truth. Subsequently, XAI methods and feature map visualization were used to interpret the internal mechanisms of the best performing model. The parameters used for model training are given in Table 2. Both models were trained using 48 different combinations of batch sizes, number of filters, and learning rates listed in the table. For the simple multiscale U-Net, the best performance was given by 32 filters, 32 batch size, and a learning rate of 1e-3 (Simple Multiscale U-Net32,32,1e−3), whereas the combination of 16 filters, 8 batch size, and 1e-3 learning rate gave attention multiscale U-Net its best result (Attention Multiscale U-Net16,8,1e−3). Both models were trained for 200 epochs using the adam optimizer and focal tversky loss function. Table 2Hyperparameters used for model training.ParameterValueFilters8, 16, 32Batch size8, 16, 32, 64Learning rate1e-3, 1e-4, 5e-4, 5e-5Epochs200OptimizerAdamLoss functionFocal tversky Table 2Hyperparameters used for model training.ParameterValueFilters8, 16, 32Batch size8, 16, 32, 64Learning rate1e-3, 1e-4, 5e-4, 5e-5Epochs200OptimizerAdamLoss functionFocal tversky Hyperparameters used for model training. Results Quantitative evaluation Table 3 presents the performance metrics of both models. Accordingly, both models achieved similar performances with the attention multiscale U-Net16,8,1e−3 having slightly higher recall (73.79%), F1 Score (81.69%), IoU (69.05%), and MCC (81.03%) compared to that of simple multiscale U-Net. These results are in the range of those reported in previous studies, as shown in Table 4, which presents the F1 Score of models developed for similar tasks. Although these differences are not directly comparable due to the differences in data, inventories, and criteria used in these studies, they can still be used as a reference to indicate that our model achieved strong performance in landslide segmentation despite relying on a relatively small number of study area samples. Table 3Performance of both models on the test dataset. The models along with the hyperparameters used for their training are given as: model Filters, batch Size, learning rate.ModelPrecisionRecallF1IoUMCCAttention multiscale U-Net16,8,1e−391.4973.7981.6969.0581.03Simple multiscale U-Net32,32,1e−392.7472.5281.0568.2380.48 Table 3Performance of both models on the test dataset. The models along with the hyperparameters used for their training are given as: model Filters, batch Size, learning rate.ModelPrecisionRecallF1IoUMCCAttention multiscale U-Net16,8,1e−391.4973.7981.6969.0581.03Simple multiscale U-Net32,32,1e−392.7472.5281.0568.2380.48 Performance of both models on the test dataset. The models along with the hyperparameters used for their training are given as: model Filters, batch Size, learning rate. Qualitative evaluation To visually analyze the performance of the models, outputs from both models were plotted alongside their corresponding images and ground truth masks for comparison. Figure 8 presents these plots. Attention multiscale U-Net made predictions that had higher similarity to the landslide mask compared to that from simple multiscale U-Net, with the latter model showing a higher number of false negatives in its output (misclassifying landslide pixels as non-landslide). This plot demonstrates the effectiveness of attention multiscale U-Net in accurately segmenting landslides while avoiding false predictions. Model interpretation using explainable AI After training and evaluating the attention multiscale U-Net model for landslide segmentation, XAI techniques namely Integrated Gradients (IG) and Gradient-Weighted Class Activation Mapping (Grad-CAM), along with feature map visualization were employed for further understanding of model performance and the relation between model inputs and output. Table 4Comparison of F1 score with selected previous studies.Model & studyF1 ScoreAttention multiscale U-Net (Current study)81.69%GELANc + CBAM (Chandra, Naveen, et al., 2024)5481.5%CResU-Net (Pham, Thang M., et al., 2024)3176.1%U-Net w/ResNet18 Backbone (Chen, Tzu-Hsin Karen, et al., 2024)2283.5 ± 3%U-Net+++ (Hu, Zekun, et al., 2023)1582%U-Net (Bhuyan, Kushanav, et al., 2023)4379.5%ADSMS U-Net (Bhuyan, Kushanav, et al., 2023)3482.33%U-Net (Meena, Sansar Raj, et al., 2023)2178.35%DeepLabV3+ (Morales, Bastian, et al., 2022)1879%Matrix SegNet (Yu, Bo, et al., 2021)4176.11%ResU-Net (Qi, Wenwen, et al., 2020)1689% Table 4Comparison of F1 score with selected previous studies.Model & studyF1 ScoreAttention multiscale U-Net (Current study)81.69%GELANc + CBAM (Chandra, Naveen, et al., 2024)5481.5%CResU-Net (Pham, Thang M., et al., 2024)3176.1%U-Net w/ResNet18 Backbone (Chen, Tzu-Hsin Karen, et al., 2024)2283.5 ± 3%U-Net+++ (Hu, Zekun, et al., 2023)1582%U-Net (Bhuyan, Kushanav, et al., 2023)4379.5%ADSMS U-Net (Bhuyan, Kushanav, et al., 2023)3482.33%U-Net (Meena, Sansar Raj, et al., 2023)2178.35%DeepLabV3+ (Morales, Bastian, et al., 2022)1879%Matrix SegNet (Yu, Bo, et al., 2021)4176.11%ResU-Net (Qi, Wenwen, et al., 2020)1689% Comparison of F1 score with selected previous studies. Integrated gradient analysis IG is an attribution method used to identify which input features contribute the most to the predictions of a model. In this case, IG identifies the contribution of each input band to the U-Net in the segmentation of landslides. An IG map, which is generated using the prediction of the model and the corresponding input, highlights the spatial regions most influential to the model’s decision for the target class (landslides). Out of all the testing samples, those with zero attribution are excluded and band-wise influence was computed for the remaining samples by performing element-wise multiplication between each band and the IG map, followed by summing the absolute values to obtain influence scores. Fig. 8Input images of landslides with their corresponding ground truth masks and predictions from each model. Yellow colour indicates landslide pixels and violet colour indicates non-landslide pixels. Input images of landslides with their corresponding ground truth masks and predictions from each model. Yellow colour indicates landslide pixels and violet colour indicates non-landslide pixels. These scores were then normalized, with the resulting values representing relative importance of each band in the prediction of the model. Among the 672 test samples analyzed, only 336 yielded non-zero IG attribution maps. Figure 9 illustrates the mean and median influence of each band across these valid samples, while Fig. 10 presents the sample-wise contribution of each spectral band to the model’s predictions. Fig. 9Bar chart depicting mean and median influence of each spectral band on the model output. Bar chart depicting mean and median influence of each spectral band on the model output. Fig. 10Line chart showing sample-wise influence of each band on the model output. Line chart showing sample-wise influence of each band on the model output. The NIR band exhibited the highest mean and median influences close to 30%. In comparison, the blue, green, and red bands each showed approximately 20% influence. This indicates that the model relied more heavily on the NIR band for landslide segmentation decisions. The sample-wise influence also clearly indicates that the NIR band consistently exerted the greatest influence on the U-Net model’s outputs, with its contribution ranging approximately from 25 to 30% in the lower end to about 45 to 50% at its peak. On the contrary, the blue, green, and red bands each had a relatively lower influence, generally falling in 20%. Notably, NIR band’s influence varied considerably across the samples, indicating a higher but more variable contribution. In comparison, the visible bands showed a more stable and consistent influence across the dataset. The higher importance of the NIR band in landslide segmentation can be attributed to its unique spectral properties, which capture critical information not available in the visible spectrum. The NIR band is particularly sensitive to vegetation health and density, making it valuable in detecting changes in forested regions such as the Western Ghats, where landslides frequently occur. Additionally, NIR is also responsive to moisture content. The changes in soil moisture resulting from landslides makes its utility higher for accurate detection. These unique properties allow the NIR band to capture vital information absent in the visible spectrum, thereby enhancing its contribution to the model’s predictions. Grad-CAM analysis Grad-CAM method gives saliency maps that highlight the regions in an input image that are the most influential for its prediction from the model. This can be analyzed for each hidden layer in the model, allowing us to understand what part of an image they focus to predict landslides from it. First, a specific convolution layer is selected from the model, and a new model is constructed to output both the activations of this layer and the final prediction. Subsequently, the gradient of the model’s prediction for the target class (in this case, landslides) is computed with respect to the activations of the chosen layer. These gradients are globally averaged across spatial dimensions to obtain importance weights for each feature map channel. Each feature map is then scaled by its corresponding weight and the weighted maps are summed to produce a coarse saliency map. Finally, the saliency map is normalized and thresholded to retain only positive values, highlighting the regions in the input that contributed most significantly to the class prediction. In this study, Grad-CAM saliency maps are generated for each of the 3 × 3 convolution layers within the U-Net model. The architecture comprises 9 major blocks; 4 encoder blocks, 4 decoder blocks, and 1 bottleneck block with each block containing 2 such convolution layers. Selected saliency maps from each block along with the sample input image and corresponding model prediction, are displayed in Fig. 11. Fig. 11Grad-CAM saliency maps for the selected input image. Conv ‘n’ shows the saliency map from the nth convolution layer in the model. Warmer colours (red, orange) indicate higher activation and contribution for image segmentation while cooler colours (blue) indicate lower activation. Grad-CAM saliency maps for the selected input image. Conv ‘n’ shows the saliency map from the nth convolution layer in the model. Warmer colours (red, orange) indicate higher activation and contribution for image segmentation while cooler colours (blue) indicate lower activation. The saliency maps of the 1 st convolution layer show considerable noise with a distinct landslide pattern. Between the layers 14 and 24, the saliency maps display minimal detail. This is attributed to the inherent nature of the convolution process, which tends to spread and dilute feature representations across neighboring regions. These dispersed features are progressively aggregated in the deeper layers, enhancing the model’s ability to highlight larger and more meaningful regions such as landslide zones as observed from 41 st layer onwards. Finally, the saliency maps from the layers 107 and 120, which are from the final block of the network, appear smoother and exhibit strong activation in the landslide regions. Although some weak activations are observed in the surrounding regions, they are suppressed by the final activation function before the binary landslide mask is generated. Feature map analysis For understanding the features learned by the model, feature maps from each filter in the model can be visualized. Since there are a large number of feature maps in the model, an interactive dashboard was implemented. Figure 12 presents a screenshot of the dashboard interface designed for this study. It includes a dropdown menu for selecting an input image from a predefined list of samples, along with sliders to navigate through the model’s layers and filters. Users can view feature maps of the selected input image from any convolution layer, both before and after the application of ReLU activation function. The sliders allow for inspection of individual filters within each layer, providing a detailed view of how a model processes and transforms the input at various stages. For reference and context, the input image, its corresponding ground truth landslide mask, and the U-Net model’s prediction are displayed alongside their respective feature maps. This tool offers a valuable means to explore the changes undergone by input layers inside the model at different depths and hidden layers. For this study, feature map visualization will be used to understand the features learned by the model in different blocks. Figure 13 shows selected feature maps from the final layers of each convolution block. This figure also illustrates how the initial encoder blocks capture scattered details from the image that are not relevant to landslide segmentation, which then gradually diminish as it goes to deeper layers and gives way to more refined and smoother feature representations that enables the model to effectively localize and segment landslide regions confidently. In the saliency maps corresponding to layers from decoder 2 to 4, a distinct grid-like pattern emerges, characterized by prominent vertical and horizontal lines. This artifact is likely the result of the upsampling process, which involves the duplication and interpolation of values across a regular grid structure. Additionally, this effect may be amplified by the integration of low-level details from the encoder through skip connections. Discussions Both models deployed in this study demonstrated strong performance in segmenting landslides, with the attention multiscale U-Net having slightly higher performance compared to the simple multiscale U-Net. This was especially evident from qualitative visual assessments, where the attention multiscale U-Net demonstrated superior performance in minimizing false negatives. The model can also be effective in other areas of Western Ghats and similar forest covered mountainous regions. However, the effectiveness of this model in other geographic regions should be figured out through further experiments since there are chances for misclassifications in different terrains such as snow covered mountains or barren ranges. Fig. 12Dashboard designed for feature map visualization. Dashboard designed for feature map visualization. Fig. 13Selected feature maps from the final layer of each convolution block in the model. Selected feature maps from the final layer of each convolution block in the model. Incorporating auxiliary datasets such as the Normalized Difference Vegetation Index (NDVI), slope, elevation, and Synthetic Aperture Radar (SAR) imagery can further enhance model performance by providing additional context and discriminatory features. However, a limitation of transfer learning is that the number and type of input bands of the pre-trained model must be consistent and closely match the target dataset. Therefore, if the source dataset used for pre-training lacks auxiliary bands, they cannot be introduced during the transfer learning phase. Since the HR-GLDD dataset used for initial training in this study only had four bands, additional bands or datasets could not be incorporated to the training data. The study used a relatively finite number of samples (222 landslide and 222 non-landslide samples) from the study area, which limits the generalization and uncertainty of the model while still showing promising performance due to transfer learning. Cross-regional validation was also not conducted since it is outside the scope of this study. Since disaster-related datasets cannot be generated at will, the reliance on publicly available data, combined with a small number of local samples, reflects the practical reality of landslide studies but also limits large-scale generalization. While the study relies on HR-GLDD for pretraining and a limited number of local samples for fine-tuning, this mirrors the real-world data scarcity in landslide research. We see this not just as a constraint but as an opportunity to demonstrate how transfer learning can bridge the gap between global datasets and local disaster contexts. Future work will focus on developing a dataset with a sufficiently large number of samples to enable model training without relying on pretrained models. This will also allow the incorporation of additional auxiliary variables as well as the use of multi-encoder architectures to process them, enhancing the model’s learning capacity and overall performance. Cross-regional validation techniques will also be adopted in future work to ensure robustness and reduce uncertainty in model performance. The use of temporal differencing between pre-event and post-event images proved beneficial in highlighting landslides by suppressing features common to both timeframes. This helped the model to focus on changes specifically caused by landslides. However, this approach introduces certain limitations. Temporal differencing may also emphasize features unrelated to landslides, such as newly emerged objects or disturbances in the post-event image. For instance, cloud cover and haze present in the post-event image can be mistakenly identified as landslides due to their absence in the pre-event image. Temporal differencing, though previously employed in other works, continues to be valuable in highlighting landslide-induced changes. Another challenge arises in areas with water bodies, which are inherently dynamic in nature. Fluctuations in water levels can expose sandbanks or sediment deposits, particularly when water levels are high in the pre-event image and low in the post-event image. These exposed features may appear as new elements in the temporal difference image and could be misclassified as landslides. To mitigate such misclassifications, careful selection of pre- and post-event imagery is crucial, ensuring that other types of land changes are minimized. Additionally, applying a waterbody mask over the area of interest can help the model ignore regions prone to such variability. The XAI techniques provided visual aid in understanding the internal mechanisms of the model. IG analysis revealed that the NIR band held the highest importance among the input features. This is likely due to the unique spatial features captured by it compared to the correlated information present in the visible bands. Additionally, the NIR band’s sensitivity to vegetation health, density, and moisture further reinforces its critical role in the model’s decision making process. Grad-CAM analysis and feature map visualization highlighted the hierarchical nature of learning in U-Net models. The initial layers captured fine details such as edges and textures, which were progressively diluted and aggregated through the convolution operations. This allows the deeper layers to focus on broader, high-level spatial patterns relevant to landslide identification. By the final convolution block, the model showed precise localization of landslide-affected areas, indicating its capability to integrate multi-scale features for accurate segmentation. Grid-like artifacts were observed in the decoder blocks, likely resulting from the upsampling operations and the incorporation of encoder features via skip connections. However, these artifacts were progressively refined and smoothed in the later stages of the network. The explainable AI methods used in this work are established techniques; however, their integration in a landslide-specific context allows us to connect model behavior with geoscientific interpretation, which is crucial for disaster risk management. Future extensions will explore tailoring interpretability tools specifically for disaster mapping tasks, thereby deepening the synergy between AI outputs and domain knowledge. Ultimately, this combination of interpretability and adaptability positions XAI-enabled models as practical tools for supporting early warning systems and resilient disaster management strategies. Conclusion This study was a first step towards the creation of an automated landslide inventory for the Western Ghats region of Kerala. We selected a small part of this region as the study area and developed a dataset from there for training and deploying two multiscale U-Net models. The attention multiscale U-Net designed in this study demonstrates promising capabilities for landslide detection as per the quantitative and qualitative evaluation. However, careful preparation of a much larger dataset, use of auxiliary variables, and advanced segmentation architectures are essential to further improve accuracy and reduce misclassifications, which can be effective in rapid and accurate mapping of landslides in such a complex and dynamic environment. The XAI techniques uncovered how the U-Net architecture interprets and processes spatial features, thereby enhancing the transparency of the landslide segmentation model. Incorporating auxiliary and terrain information such as DEM, NDVI, and SAR and applying XAI methods on these inputs can further strengthen explainability by revealing which variables most influence the model’s decisions and under what conditions. Such an approach not only advances our understanding of how deep learning models process heterogeneous geospatial data across different layers but also offers actionable knowledge that can guide practitioners in dataset preparation, model selection, and operational deployment in diverse terrains.