Impact of geological map scale on the performance of landslide susceptibility zonation using artificial intelligence models

Abstract

This study assesses the impact of geological map scale on the performance of artificial intelligence (AI) models in landslide susceptibility zonation (LSZ) along the Tuyen Quang–Ha Giang Highway, northern Vietnam. For this purpose, a total of 139 past landslide locations and 14 conditioning factors were collected and used for generating training (70%) and testing (30%) datasets. Four AI models: Artificial Neural Network (ANN), Convolutional Neural Network (CNN), Deep Neural Network (DNN) and a hybrid CNN–DNN model were developed and used for landslide modeling at two geological map scales: a small-scale map (1:200000) and a large-scale map (1:50000). Model accuracy was evaluated using AUC/ROC and multiple statistical indices. Results of this study showed that all models demonstrated effective predictive performance (AUC = 0.89–0.97), with the hybrid CNN–DNN model achieving the best performance and generalization. The models using the larger-scale geological map yielded higher spatial precision and better delineation of high and very high susceptibility zones compared to those used the smaller-scale geological map. The SHAP analysis revealed that maximum rainfall, average annual rainfall, vegetation cover (NDVI), elevation, and proximity to roads were the most influential factors to the predictive capability of the models. The study also highlights the importance of AI-based models, particularly deep learning and hybrid architectures in capturing complex nonlinear relationships among conditioning factors, improving prediction accuracy, and enhancing the reliability of LSZ for hazard management and infrastructure planning in mountainous regions.