Landslide susceptibility mapping using deep learning in Dak Lak, Vietnam

Abstract

Landslides cause significant damage to transportation infrastructure, settlements, and agricultural land in the Central Highlands of Vietnam, particularly in the former Dak Lak province. Therefore, identifying areas with the potential for landslide occurrence is essential to support land-use planning and the development of landslide mitigation measures. This study proposes an explainable deep-learning framework for landslide susceptibility assessment using four models: Convolutional Neural Network (CNN), Deep Residual Dense Network (DRDN), Artificial Neural Network (ANN), and Kolmogorov–Arnold Network (KAN). An inventory comprising 235 landslide polygons was analyzed together with 14 conditioning factors, including aspect, slope, curvature, elevation, geology, land use, soil type, rainfall, distance to faults, distance to roads, distance to rivers, the normalized difference vegetation index (NDVI), the topographic wetness index (TWI), and flow accumulation. The results showed that CNN achieved the best performance, with an area under the curve (AUC) = 0.913, a mean absolute error (MAE) = 0.197, and a root mean square error (RMSE) = 0.444 for the training dataset, and AUC = 0.888, MAE = 0.251, and RMSE = 0.501 for the validation dataset. Model interpretation showed that landslide susceptibility was notably influenced by slope, mean annual rainfall, geological factors, and distance to roads. The effects of these factors were nonlinear and varied across different value ranges, while their response patterns also differed among individual locations. A spatial accuracy assessment of the susceptibility map showed that the high and very high susceptibility classes occupied only 18.30% of the study area but contained 94.88% of the observed landslide pixels. In particular, the very high susceptibility class occupied 7.51% of the study area, contained 68.18% of the landslide pixels, and recorded the highest frequency ratio of 9.079. The results demonstrate that explainable deep-learning models can not only support the identification of areas with potential for landslide occurrence but also clarify the roles and response patterns of the conditioning factors, thereby providing a basis for land-use planning and the development of landslide mitigation measures in the study area.