The ClearSCD model: Comprehensively leveraging semantics and change relationships for semantic change detection in high spatial resolution remote sensing imagery

• Published in: ISPRS journal of photogrammetry and remote sensing 2024-05, Vol.211, p.299-317
• Main Authors: Tang, Kai, Xu, Fei, Chen, Xuehong, Dong, Qi, Yuan, Yuheng, Chen, Jin
• Format: Article
• Language: English
• Subjects: Change detection; Deep learning; Multi-task learning; Remote sensing; Semantic change detection; Semantic segmentation
• ISSN: 0924-2716; 1872-8235
• DOI: 10.1016/j.isprsjprs.2024.04.013

•ClearSCD considers the bidirectional dependency between semantics and change.•A BSCC mechanism establishes the correlation of bi-temporal semantic features.•A deep CVAPS module enhances the efficiency of change detection.•A SACL module optimizes feature space of semantics explicitly.•A new semantic change detection dataset, LsSCD, was proposed by this study. The Earth has been undergoing continuous anthropogenic and natural change. High spatial resolution (HSR) remote sensing imagery provides a unique opportunity to accurately reveal these changes on a planetary scale. Semantic change detection (SCD) with HSR imagery has become a common technique for tracking the evolution of land surface types at a semantic level. However, existing SCD methods rarely model the dependency between semantics and changes, resulting in suboptimal accuracy in detecting complicated surface changes. To address this limitation, we propose ClearSCD, a multi-task learning model that leverages the mutual gain relationship between semantics and change through three innovative modules. The first module interprets semantic features at different times into posterior probabilities for surface types to detect binary change information; the second module learns the correlation between surface types over time and the binary change information; a semantic augmented contrastive learning module is used as the third module to improve the performance of the other two modules. We tested ClearSCD’s performance against state-of-the-art methods on benchmark datasets and a real-world scenario (named LsSCD dataset), showing that ClearSCD outperformed the alternatives on mIoUsc metrics by 1.23% to 19.34%. Furthermore, ablation experiments demonstrated the unique contribution of the three innovative modules to performance improvement. The high computational efficiency and robust performance over diverse landscapes demonstrate that ClearSCD is an operational tool for detecting detailed land surface changes from HSR imagery. Code and LsSCD dataset available at https://github.com/tangkai-RS/ClearSCD.