Dual-stream feature fusion network for landslide detection

Landslides are destructive natural hazards that threaten infrastructure, ecosystems, and human settlements, particularly in mountainous regions. Although deep learning has significantly improved landslide detection from remote sensing imagery, existing models often struggle with complex terrain conditions or require high computational resources. To address these challenges, this study proposes a Dual-Stream Feature Fusion Network (DSFF-Net) for accurate and efficient landslide segmentation. The proposed architecture integrates dual-stream feature extraction, hierarchical feature fusion, and an Enhanced Spatial Attention (ESA) module to effectively capture both spatial details and high-level semantic information. DSFF-Net was evaluated using two benchmark datasets, Bijie and Landslide4Sense. Experimental results show that the model achieves 96.34% precision, 97.83% recall, 97.08% F1-score, and 94.32% mIoU on the Bijie dataset, and 77.05% precision, 64.48% recall, 70.20% F1-score, and 72.36% mIoU on the Landslide4Sense dataset. Despite its compact architecture, DSFF-Net maintains low computational complexity with approximately 24.51 million parameters, demonstrating a strong balance between segmentation accuracy and computational efficiency. These results indicate that the proposed framework provides a reliable approach for automated landslide detection from remote sensing imagery and may support large-scale landslide inventory mapping and environmental monitoring applications. Future studies should further evaluate the model across diverse geographic regions and integrate multi-source geospatial data to enhance its generalization and operational applicability for landslide hazard monitoring.