ConFormer-Net: Spatiotemporal Modeling for Landslide Detection Using Multi-Temporal SAR Data

Abstract

Highlights What are the main findings? ConFormer-Net combines multiscale dilated convolutions, sinusoidal positional encoding, Transformer-based temporal modeling, and cross-attention to jointly capture spatial structures and long-range temporal changes in multi-temporal SAR imagery. ConFormer-Net achieved an F1-score of 95.27%, an accuracy of 95.29%, a precision of 96.79%, and a recall of 93.79%, achieving the best overall performance among all comparison models; ablation experiments identified temporal modeling as the primary contributor to performance improvement. What are the implications of the main findings? Joint spatiotemporal modeling improves the discrimination of landslide-induced changes from vegetation disturbance, terrain shadows, bare ground, and SAR speckle noise, thereby reducing false detections in complex mountainous environments. ConFormer-Net achieved high detection accuracy and stable performance under different terrain conditions and landslide morphologies, providing an effective approach for landslide detection in complex terrain environments. Abstract Landslides are characterized by sudden occurrence, severe destructiveness, and widespread spatial distribution. Therefore, rapid and accurate landslide detection is essential for reducing infrastructure damage and safeguarding human lives. Conventional landslide monitoring methods are generally time-consuming, labor-intensive, and inefficient, while existing deep learning-based landslide detection methods remain limited in multiscale spatial structure representation, temporal sequence modeling, and spatiotemporal feature fusion. To address these limitations, this study proposes ConFormer-Net, a spatiotemporal landslide detection model that integrates convolutional neural networks (CNNs) with a Transformer architecture for landslide detection from multi-temporal synthetic aperture radar (SAR) imagery. The proposed model adopts a hybrid architecture and incorporates a cross-attention mechanism. Specifically, a dilated convolutional network is employed to extract local spatial features of landslides, while a Transformer encoding module models the temporal dependencies among multi-temporal SAR observations. The extracted spatial and temporal features are subsequently integrated through the cross-attention mechanism to achieve accurate landslide detection. The overall detection performance of ConFormer-Net was first evaluated using samples from different regions in the publicly available Sen12Landslides dataset. The proposed model was then compared with CNN, CNN-LSTM, ConvLSTM, GRU, CNN3D and ResNet50 models. The results demonstrate that ConFormer-Net achieved the best overall performance, with an F1-score of 95.27%, an accuracy of 95.29%, a precision of 96.79%, and a recall of 93.79%. These results indicate that ConFormer-Net enables highly accurate landslide detection while maintaining moderate model complexity, demonstrating its effectiveness for landslide detection from multi-temporal SAR imagery.