Enhancing Semantic Segmentation-based Change Detection in Bi-temporal Satellite Images

Abstract

Remote sensing satellite imagery has become an important tool in change detection and is essential for monitoring the environment, urban development, and land use. Conventional approaches often fail to reflect the subtle geospatial and temporal changes that comprise dynamic landscapes. The availability of satellite and aerial imagery has revealed the need for sophisticated algorithms. Recent deep learning developments have potential for satellite image analysis because they can extract features and recognize patterns. The proposed early fusion framework incorporates a UNet++ architecture with a Reg- NetY_016 backbone, reinforced with SCSE attention mechanisms in the decoder block. By learning hierarchical features, the proposed model effectively captures fine spatial details and subtle variations. Comprehensive testing on Szada/1 and Tiszadob/3, on standard splits used for comparison with existing state-of-the-art methods, achieves F1-scores of 75.54% and 92.68%, respectively. The method also outperforms with an average F1-score of 87.87% ± 4.30% (Szada) and 93.72% ± 1.55% (Tiszadob) under 5- fold cross-validation on the full datasets across different data splits. Ablation experiments and computational complexity analysis validate that both nested skip connections and the SCSE attention module boost the detection performance with minimal computational overhead. Cross-region experiments further confirm the proposed model’s transferability across distinct datasets. These results show that deep learning architectures coupled with specialized attention mechanisms are viable approaches for robust performance on change detection tasks in bi-temporal satellite images.