MSFF-CDNet: A Multiscale Feature Fusion Change Detection Network for Bi-Temporal High-Resolution Remote Sensing Image
Lukang Wang, Yue Li, Min Zhang, Xiaoqi Shen, W. Peng, Wenzhong Shi
Abstract
Change detection (CD) is an important application of remote sensing (RS) technology, which discovers changes by observing bi-temporal RS images. The rise of deep learning provides new solutions for CD. However, due to the insufficient extraction and utilization of deep features from RS images, existing deep learning-based CD methods are difficult to fully integrate such deep features, resulting in unstable performance, especially low sensitivity to multi-scale changes. In this letter, a multi-scale feature fusion CD network (MSFF-CDNet) is proposed to enhance feature fusion, by integrating a mask guided change fusion module (MGCF) to achieve the fusion of the consistency and difference of multi-scale features. Also, a CD refinement module (CDRM) is implemented to assist the encoding-decoding structure to achieve CD at a finer scale. By training with a hybrid loss function, the MSFF-CDNet is able to learn trans-formation relationships of bi-temporal RS images and their change maps. Besides, using a deep supervised learning strategy further improves the fitting performance and robustness. The method is experimented on two open-source datasets (i.e., CDD and LEVIR-CD dataset). Compared to state-of-the-art CD methods, the proposed method outperforms on all metrics and its IoU reaches 92.39% and 85.89%, respectively. The codes are available at https://github.com/WangLukang/MSCD.