BGSNet: A boundary-guided Siamese multitask network for semantic change detection from high-resolution remote sensing images

Abstract

The accurate identification of land-surface changed classes when locating changed areas with regular boundaries from satellite images represents a significant challenge, particularly in these areas with considerable spectra and seasonal differences. This paper develops a boundary-guided Siamese multitask network, namely BGSNet, for the purpose of semantic change detection (SCD) from high-resolution remote sensing images. The objective of BGSNet is to utilize robust boundary semantics to enhance the intra-class consistency of change features, alleviating the pseudo-changes caused by temporal variances while retaining well boundary details. The proposed BGSNet consists of three tasks including bi-temporal semantic segmentation, changed areas detection, and boundary detection tasks. In particular, for the semantic segmentation task, a Siamese multilevel pyramid network based on transformer as feature extractors is introduced to fully capture robust semantic features of bi-temporal remote sensing images. For the boundary detection task, a multi-scale feature decoder is designed to enhance boundary semantic representation. For the change detection task, a boundary-contextual guided module is constructed to supply fine-grained semantic constraints, refining the boundaries of detected areas. Finally, we introduce a multitask self-adaptive weighting loss function that considers task uncertainty, which effectively balances the learning effects of different tasks, and improves the model’s adaptability to varied semantic change scenarios. Extensive experiments were conducted on the JiLin-1 and HRSCD datasets, demonstrating that BGSNet outperformed the eight state-of-the-art methods in identifying various semantic changes. Our methods produced the highest attribute accuracy, exceeding reference methods by 5.41%-29.63% and 2.94%-28.09% in SeK measures. Moreover, the detected results by BGSNet exhibited excellent boundary consistency with ground truth, resulting in the lowest geometric errors GTC of 0.187, 0.308, and 0.321 on the JiLin-1, HRSCD, and Fuzhou datasets, respectively. The proposed method also showed a high application promise in large-scale cropland non-agriculturalization scenarios with significant temporal and spectral variations. The code and data will be made available at https://github.com/long123524/BGSNet.