ISBN: 978-981-18-1791-5 DOI: 10.18178/wcse.2021.06.012
LSDNet: Boosting Change Detection of High-resolution Remote Sensing Images by Combining Convolution-Involution and Ensemble Coordinate Attention
Abstract— Change detection has always been a crucial task in remote sensing fields, and there have already been great efforts made on it for decades. However, as high resolution (HR) remote sensing images generally contain abundant ground details, it still faces a huge challenge for their change detection, especially from the aspects of change detection accuracy and speed. Concerning this issue, a novel lightweight Siamese deep network (LSDNet) is proposed, and it combines Convolution-Involution Module (CIM) and Ensemble Coordinate Attention Module (ECAM) for boosting the change detection of HR remote sensing images. CIM summarizes the context of ground objects and reweights the importance of different positions, while ECAM aggregates multiple levels of semantic features and pays different attention to different spatial information. The experiments on CNZ data set have shown that the proposed LSDNet performs better than state-of-the-art (SOTA) change detection methods, especially it improves the accuracy by 1.92% and reduces the amount of model parameters by 32.89% compared to SNUNet-CD which has the best performance currently.
Index Terms— artificial intelligence, change detection, Ensemble Coordinate Attention, high-resolution remote sensing images
Yifan Liu, Jingdong Liu, Asif Raza, Zeng Li, Hong Huo, Tao Fang
Department of Automation, Shanghai Jiao Tong University, CHINA
Key Laboratory of System Control and Information Processing, Ministry of Education of China, CHINA
Shanghai Engineering Research Center of Intelligent Control and Management, CHINA
Cite: Yifan Liu, Jingdong Liu, Asif Raza, Zeng Li, Hong Huo, Tao Fang , "LSDNet: Boosting Change Detection of High-resolution Remote Sensing Images by Combining Convolution-Involution and Ensemble Coordinate Attention ," 2021 The 11th International Workshop on Computer Science and Engineering (WCSE 2021), pp. 79-84, Shanghai, China, June 19-21, 2021.