Enhanced Landslide Detection Using a Swin Transformer With Multiscale Feature Fusion and Local Information Aggregation Modules

In recent years, detecting and monitoring landslides have become increasingly critical for disaster management and mitigation efforts. Here, we propose a model for landslide detection utilizing a combination of the Swin Transformer architecture with multiscale feature fusion lateral connection and local information aggregation modules (LIAMs). The Swin Transformer, known for its effectiveness in image understanding tasks, serves as the backbone of our detection system. By leveraging its hierarchical self-attention mechanism, the Swin Transformer can effectively capture both local and global contextual information from input images, facilitating accurate feature representation. To increase the performance of the Swin Transformer specifically for landslide detection, we introduce two additional modules: the multiscale feature fusion lateral connection module (MFFLCM) and the LIAM. The former module enables the integration of features across multiple scales, allowing the model to capture both fine-grained details and broader contextual information relevant to landslide characteristics. Meanwhile, the latter module focuses on aggregating local information within regions of interest, further refining the model’s ability to discriminate between landslide and non-landslide areas. Through extensive test and evaluation of benchmark datasets, our proposed method demonstrates promising results in detecting landslides with high mIoU, $F1$ score, kappa, precision, and recall 84.2%, 90.7%, 82.6%, 89.9%, and 91.9%, respectively. Moreover, its robustness to variations in terrain and environmental conditions suggests its potential for real-world applications in landslide monitoring and early warning systems. Overall, our study highlights the effectiveness of integrating advanced transformer architectures with tailored modules for addressing complex geospatial challenges like landslide detection.