TransSiamUNet Based Transformer-Augmented Siamese-U-Net for Precise Change Detection in Satellite Imagery

Abstract

Identifying changes in satellite images is vital for tasks like tracking land cover and land use, evaluating disaster impacts, and conducting military surveillance. Although conventional techniques for detecting changes in multispectral remote sensing data are commonly applied, they often fail to meet the requirements for reliability and precision. Recently, deep learning methods have emerged, providing more accurate and effective solutions for monitoring environmental transformations and urban expansion in satellite imagery. This paper introduces TransSiamUNet, a deep learning architecture that combines Siamese networks, U-Net segmentation, and Vision Transformers (ViT) for high-precision change detection. The model processes paired Sentinel-2 images via a tailored preprocessing pipeline and integrates local and global feature extraction for pixel-level change segmentation. On the OSCD benchmark, TransSiamUNet achieves an accuracy of 0.94, surpassing the Siamese network (0.86), U-Net (0.84), and Siamese+U-Net hybrid (0.91). These results demonstrate the model’s superior capability in detecting fine-grained urban and environmental changes, highlighting its suitability for real-world remote sensing applications.