MLDFNet: A Multilabel Dual-Flow Network for Change Detection in Bitemporal Remote Sensing Images

Abstract

With the development of deep learning (DL) in recent years, numerous remote sensing image change detection (CD) networks have emerged. However, existing DL-based CD networks still face two significant issues: 1) the lack of adequate supervision during the encoding process; and 2) the coupling of overall information with edge information. To overcome these challenges, we propose the Edge detection-guided (ED-guided) strategy and the Dual-flow strategy, integrating them into a novel Multilabel Dual-flow Network (MLDFNet). The ED-guided strategy supervises the encoding process with our self-generated edge labels, enabling feature extraction with reduced noise and more precise semantics. Concurrently, the Dual-flow strategy allows the network to process overall and edge information separately, reducing the interference between the two and enabling the network to observe both simultaneously. These strategies are effectively integrated through our proposed Dual-flow Convolution Block. Extensive experiments demonstrate that MLDFNet significantly outperforms existing state-of-the-art methods, achieving outstanding F1 scores of 91.72%, 97.84%, and 94.85% on the LEVIR-CD, CDD, and BCDD datasets, respectively. These results validate its superior performance and potential value in real-world remote sensing applications.