Analysis of the impact of terrain factors and data fusion methods on uncertainty in intelligent landslide detection

Abstract

Current research on deep learning-based intelligent landslide detection modeling has focused primarily on improving and innovating model structures. However, the impact of terrain factors and data fusion methods on the prediction accuracy of models remains underexplored. To clarify the contribution of terrain information to landslide detection modeling, 1022 landslide samples compiled from Planet remote sensing images and DEM data in the Sichuan–Tibet area. We investigate the impact of digital elevation models (DEMs), remote sensing image fusion, and feature fusion techniques on the landslide prediction accuracy of models. First, we analyze the role of DEM data in landslide modeling using models such as Fast_SCNN, the SegFormer, and the Swin Transformer. Next, we use a dual-branch network for feature fusion to assess different data fusion methods. We then conduct both quantitative and qualitative analyses of the modeling uncertainty, including examining the validation set accuracy, test set confusion matrices, prediction probability distributions, segmentation results, and Grad-CAM results. The findings indicate the following: (1) model predictions become more reliable when fusing DEM data with remote sensing images, enhancing the robustness of intelligent landslide detection modeling; (2) the results obtained through dual-branch network data feature fusion lead to slightly greater accuracy than those from data channel fusion; and (3) under consistent data conditions, deep convolutional neural network models and attention mechanism models show comparable capabilities in predicting landslides. These research outcomes provide valuable references and insights for deep learning-based intelligent landslide detection.