Self-structured pyramid network with parallel spatial-channel attention for change detection in VHR remote sensed imagery.

• We propose a new deep learning-based CD method, S2PNet, to combat the difficulties brought by some unique features of VHR remote sensing images. More precisely, the challenges of inadequate pattern separability and high land cover diversity are further overcome by our method when dealing with CD tasks. • We propose a novel feature pyramid module, SFP, to cope with multiscale change objects through the integration of the features at different layers with different spatial sizes. Compared to other PP-based modules, our SFP can acquire more authentic location information of multi-scale objects in VHR images. • We propose a dual-dimensional attention mechanism, PSA. It has two branches which are developed to refine feature maps in different dimensions, i.e., spatial-wise and channel-wise branches. Different with conventional similar mechanisms, the two branches of PSA run fully parallel, which will eliminate the interference with each other. • Comprehensive experiments have been conducted over several challenging public large-scale VHR change detection data sets. And corresponding experimental results indicate that the proposed S2PNet is able to outperform other state-of-the-art CD methods. Land cover change detection (CD) in very-high-resolution (VHR) images is still impeded by weak pattern separability and land cover complexity. To address these challenges, a self-structured pyramid network (S 2 PNet) with a parallel spatial-channel attention mechanism (PSAM) and a self-structured feature pyramid (SFP) is proposed for a finer annotation of changed land cover. The proposed PSAM refines the features of different levels in dual-branch coordinated by running parallel without mutual influence for a better recognition of varied objects, which can lead to less incorrectly detected land cover. And the SFP integrates the embedded multi-scale features to acquire an improved cognition over multi-scale objects, which can contribute to a more complete annotation over diverse objects. All-round experiments over several widely used open large-scale VHR CD data sets are carried out, which indicate the efficiency and effectiveness of the proposed method. Related comparisons suggest that the proposed method can achieve higher performance over several existing state-of-the-art CD methods. The source codes will be released at https://github.com/HaiXing-1998/S2PNet-CD. [ABSTRACT FROM AUTHOR]