Аннотация
The change detection of bi-temporal satellite images is a remote sensing problem that has been applied in many fields, such as urban monitoring, disaster assessment and environment analysis. Yet, correct detection is difficult because of the presence of the high class imbalance as well as the necessity of detecting the changes at various spatial scales. In this paper, the SiameseResUNet, an encoder–decoder architecture for satellite image change detection designed for recall-oriented change detection, is proposed. The proposed architecture exploits a shared-weight residual Siamese encoder, multi-scale contextual aggregation using Atrous Spatial Pyramid Pooling (ASPP) and explicitly enforces temporal difference connections, called difference-aware skip connections, between image pairs. Considering the lack of change pixels, a novel RecallBiasedLoss is introduced combining label-smoothed cross-entropy, asymmetric Tversky loss and Dice loss to highlight the detection of minority classes while providing stability in the predictions. Experimental results show that SiameseResUNet achieves the best F1-score, Recall, Precision, Cohen’s Kappa, and AUC-ROC among several representative state-of-the-art methods, such as BiasUNet, ChangeFormer and ChangeMamba, on the ONERA Satellite Change Detection (OSCD) dataset. Ablation studies validate the effectiveness of bottleneck, difference-aware skip connections, and the suggested loss formulation, achieving the maximum individual gain with RecallBiasedLoss. In addition, cross-dataset evaluation shows that the learned representations transfer to the OSCD domain with little fine-tuning when evaluated on LEVIR-CD. The results demonstrate that SiameseResUNet is an effective and efficient solution for recall-oriented satellite image change detection under extreme class imbalance.