A Dual-Stream Coding Fusion Network for Landslide Mapping Based on Multisource Remote Sensing Images

The accurate acquisition of fine-scale landslide information is critical for disaster early warning and prevention. However, existing landslide detection methods mainly focus on the spatial characteristics of multi-source remote sensing data, while often overlooking the cross-channel semantic dependencies inherent in their fusion. This limitation hinders the comprehensive characterization of landslides in complex environments, leading to insufficient recognition accuracy across different shapes and scales, and resulting in missed detections and false alarms. To address this challenge, we propose a Dual-Stream Coding Fusion Landslide Detection Network (DSCF-LDNet), which simultaneously models spatial features and cross-channel semantic dependencies in multi-source landslide remote sensing data, thereby enabling high-precision automatic landslide detection. Specifically, the spatial branch employs a multi-scale residual network to extract fine-grained spatial features, whereas the sequential branch integrates a multi-scale residual network with a long short-term memory module (MSRNet-LSTM) to model cross-channel correlations among multi-source features. Subsequently, a feature fusion module is designed to integrate both representations, generating more discriminative landslide features. In the decoding stage, an attention gating decoder module (AGDM) is designed to enhance multi-scale feature interactions, thereby improving the model’s sensitivity to landslide targets. Experiments on the self-constructed Bailong River Basin dataset demonstrate that DSCF-LDNet achieves a precision of 80.84%, recall of 77.89%, F1-score of 79.35%, and mIoU of 81.74%, surpassing benchmark models such as SCDUNet++, STUNet, and Attention-UNet. Tests on Landslide4Sense and Lanzhou Urban Area datasets further confirm the model’s robustness across diverse geological settings. Overall, DSCF-LDNet effectively learns both spatial features and inter-channel semantic dependencies from multi-source remote sensing data, enabling accurate landslide mapping for diverse shapes and scales, and provides a practical framework for large-scale landslide monitoring.