Classification of Surface Deformation Using DTW-Transformer Model in Lower Jinsha River Basin

InSAR-based monitoring and classification of surface deformation are essential for the timely detection and mitigation of geological hazards in reservoir areas. However, the intricate effects of environmental factors, the nonlinear nature of deformation delays, and the overwhelming data volumes continue to impede the establishment of universal approaches for the rapid classification of InSAR-derived time-series surface deformation in reservoir areas. In this study, we developed a method combining Dynamic Time Warping (DTW) and a Transformer to process deformation data extracted from SBAS-InSAR. The proposed method uses the similarity derived from the DTW distance matrix as the training objective for the Transformer model. During inference, the trained model generates embeddings for the entire study area through a single forward pass, without requiring further pairwise DTW computation. We also combined prototype assignment with unsupervised clustering to identify different deformation patterns. The experiment was conducted in the Lower Jinsha River Basin, where newly constructed reservoirs have triggered extensive surface deformation responses. With DTW's time-elastic alignment and Transformer's self-attention mechanisms, our model better distinguished deformation time series than static similarity methods, while maintaining linear computational complexity with respect to sample size during inference. The proposed method identified different deformation patterns in the study area and found that their spatiotemporal characteristics exhibit complex relationships with water level fluctuations, precipitation, and lithology.