STIFormer: Spatial-Temporal Interactive Transformer for Dual-Phase Remote Sensing Images Semantic Change Detection

The Semantic Change Detection (SCD) in remote sensing images requires accurate extraction of change regions while distinguishing between different types of changes. The temporal span between the two phases is often considerable, resulting in a diverse range of change types across a wide imaging field, with some changes being more apparent than others. Additionally, the scale of change regions can vary from local to global, posing significant challenges for SCD in dual-phase remote sensing images. To address these issues, we propose a semantic change detection method leveraging spatio-temporal dependencies. Firstly, we individually process the dual-phase remote sensing images using an encoder structure with shared weights to extract semantic-level features. These features are then subjected to convolutional operations to generate representations of change regions, enhancing the capability to extract less obvious changes. Subsequently, a transformer structure representing the semantic changes in the dual-phase images is introduced. This structure effectively utilizes the spatio-temporal characteristics of the images to obtain global and change features, addressing the issue of insufficient feature representation for large-scale change regions. Finally, we introduce a semantic learning scheme that utilizes spatio-temporal constraints consistent with the SCD task to guide the learning of semantic changes. The proposed network is evaluated on datasets obtained from high-resolution and resource series satellites, achieving an IOU of 94.49% and an accuracy of 99.04%. These results convincingly demonstrate the effectiveness of our proposed method.