SPATIAL SPARSITY AWARE EXPLAINABLE DEEP LEARNING-BASED LANDSLIDE SUSCEPTIBILITY MAPPING: APPLICATION TO A HILL DISTRICT, BANGLADESH

Golam Murad , Md. Aftabur Rahman, Hideaki Yasuhara

Landslide susceptibility mapping is a critical disaster risk management tool in mountainous regions, particularly in developing countries and in regions where development is ongoing or planned. This research introduces a novel approach to landslide susceptibility mapping that addresses the persistent challenge of spatial sparsity in landslide datasets, particularly in developing countries where strong monitoring of hill slopes is seldom available. A methodological framework that addresses data sparsity and explainable artificial intelligence (XAI) has been developed to enhance the accuracy of landslide susceptibility mapping, particularly in datasets that are limited or sparsely distributed. A landslide susceptibility map for Rangamati, the largest district of Bangladesh and part of the rugged, hilly Chittagong Hill Tract Region along the Bangladesh-India border, has been prepared using an updated methodological framework. The methodological framework first quantifies spatial sparsity through Voronoi-based clustering and spatial autocorrelation metrics (Getis-Ord Gi* and Moran's I), then implements sparsity mitigation techniques using DBSCAN clustering and Spatial density analysis. The research has examined twelve landslide conditioning factors derived from remote sensing data, including topographic, hydrological, and environmental variables. Three deep learning architectures—Deep Neural Network (DNN), one-dimensional Convolutional Neural Network (1D-CNN), and Long Short-Term Memory (LSTM)—have been implemented on both the original and sparsity-mitigated datasets. Comparative analysis revealed that the 1D-CNN model applied to the sparsity-free dataset achieved superior performance with an AUC of 0.9625, accuracy of 0.89, precision of 0.914, and F1-score of 0.853. SHAP analysis has provided unprecedented insights into feature importance, demonstrating that spatial context features. Despite limited data and computational constraints, this represents a significant improvement over models trained on the original dataset. The study demonstrates that addressing spatial sparsity before implementing deep learning algorithms substantially enhances landslide susceptibility mapping accuracy, providing a more reliable foundation for disaster risk reduction strategies. The framework demonstrates that addressing spatial sparsity before applying deep learning algorithms substantially improves prediction reliability, while XAI techniques ensure model transparency, which is essential for stakeholder confidence and operational implementation in disaster risk management scenarios.