ReSCD-Net: Resolution-Aware Deep Learning for Building Change Detection in High-Resolution Satellite Imagery

Building change detection (BCD) using high-resolution satellite data is critical in urban surveillance, disaster management, and environmental assessments. However, issues such as resolution inconsistencies, spectrum fluctuation, and urban heterogeneity frequently restrict the reliability of current approaches. We proposed a novel Resolution-Standardized Change Detection Network (ReSCD-Net) to address these problems. The framework features an Adaptive Resolution Normalizer (ARN) for dynamic resolution standardization, a Resolution-Aware Segmentation Core (RASC) with a Hierarchical Spatial Extractor (HSE) for semantic representation at various scales, and a Contextual Refinement Decoder (CRD) for precise boundary reconstruction. Multi-scale characteristics are further enhanced by Atrous Spatial Pyramid Pooling (ASPP), allowing for robust contextual reasoning and detailed edge preservation. In addition, a Balanced Structural Loss (BSL) function is added to improve the trade-off between precision and recall while suppressing spurious changes caused by shadows, seasonal variations, and sensor noise. Three benchmark datasets covering various architectural and geographic circumstances (EGY-BCD, LEVIR-CD+, and DSIFN-CD) were used for the experiments. Achieving an F1-score of 91.09% and a mIoU of 91.12% on the EGY-BCD dataset, ReSCD-Net outperformed recent CNN- and Transformer-based baselines. With an F1-score of 93.06% and a mIoU of 90.83% on LEVIR-CD+, the model surpassed transformer-enhanced techniques while preserving a lower computing complexity. ReSCD-Net's balanced precision and recall scores on DSIFN-CD further validated its capacity for generalization in diverse urban and rural settings.