SWINSHIFT-CD: TRACKING TERRESTRIAL TRANSFORMATION WITH SHIFTED WINDOWS

Change detection in remote sensing images is a field of Geospatial technology that requires models that can jointly capture fine-grained spatial details and long-range contextual information. Despite the success of vision transformers (ViTs) and Convolutional Neural Networks (CNNs) as backbones in numerous computer vision applications, they remain underutilised in change detection due to computational complexity of transformers and local sensitivity of CNNs. In this paper, we propose SwinShift-CD a hybrid architecture of Swin Transformer (Shifted window Transformer) integrated with ResNet18 backbone via a gated cross-attention Feature Injector. The ResNet-18-based detail-capture module extracts spatial details, while a feature injector fuses these into deep semantic layers to jointly model broad structural alterations. A U-style decoder then progressively reconstructs high-resolution change masks. Experiments on the LEVIR-CD benchmark show that SwinShiftCD achieves an F1 score of 91.02% and an IoU of 85.03% using the Swin-Base variant, while producing sharper boundaries and fewer false positives than recent baseline methods. Comprehensive quantitative and qualitative analyses further validate the architecture’s robustness, which demonstrates SwinShift-CD’s practical suitability for urban change monitoring.
Index Terms—Change Detection, Remote Sensing, Swin Transformer, Vision Transformer (ViT), Convolutional Neural Networks (CNNs), High-Resolution Aerial Imagery, Detail-Capture Module, Feature Fusion, Multiscale Reasoning, LEVIR-CD Dataset, ResNet18, Satellite Images