Which Geo-Auxiliary Information Improves Landslide Segmentation? A Systematic Study Using a Residual Dual-Branch SegFormer

Rapid post-event landslide mapping is essential for emergency response and disaster management; however, accurately delineating rainfall-induced landslides remains challenging because of their fragmented morphology, heterogeneous landscapes, and the complementary nature of Earth Observation data. Although recent deep learning approaches have demonstrated promising performance using optical imagery, the contribution of different geo-auxiliary information to landslide segmentation and its transferability across regions remain insufficiently understood. This study proposes a Residual Dual-Branch SegFormer that integrates heterogeneous geospatial information through a lightweight residual Convolutional Neural Network (CNN) encoder and a novel Multi-Scale Geo-Auxiliary Gated Fusion Module (MGFM), enabling adaptive feature fusion while preserving the original SegFormer architecture. Using a multimodal dataset constructed from the March 2023 Tropical Cyclone Freddy event in southern Malawi, we systematically evaluate RGB imagery combined with optical, terrain, and Sentinel-1 SAR-derived auxiliary layers. The framework is assessed on a geographically independent test area and further validated through zero-shot transfer on the public Landslide4Scene and Lushan datasets. Experimental results demonstrate the following: (1) Integrating pre- and post-event SAR coherence with RGB imagery achieves the best overall performance when applied to areas that match the training data, reaching a mean Intersection over Union (mIoU) of 80.56% and an F1-score of 78.03% in the independent Malawi test area. (2) RGB combined with slope represents the most effective among the evaluated static topographic configurations in zero-shot transfer to the two external datasets, consistently outperforming the RGB baseline. (3) The proposed framework exhibits strong cross-region generalization, achieving the highest transfer performance on both external datasets while demonstrating that SAR coherence provides the most discriminative auxiliary information for post-event rainfall-induced landslide segmentation. These findings provide practical guidance for designing robust multimodal deep learning frameworks for operational landslide mapping.