Temporal Information Enhanced Graph Convolutional Network With Superpixel-Pixel Gated Knowledge Dynamic Selection for Change Detection in Satellite Time Series

Abstract

Change detection (CD) in satellite time series is one of the hot research topics in the field of remote sensing (RS). Recently, graph convolutional network (GCN)-based CD methods have garnered significant attention from numerous researchers, as GCN can perform convolutional operations on irregular graphs using superpixels. However, these methods still face two challenges, i.e., inaccurate superpixel segmentation may lead to detection errors, and temporal information is ignored in the GCN-based CD methods. To this end, we propose a temporal information enhanced GCN with superpixel-pixel gated knowledge dynamic selection for CD (TSCD) in satellite time series. This method fully leverages the strengths of both GCN and convolutional neural network (CNN) frameworks to accurately identify changes in satellite time series. It is composed of three parts: First, a superpixel-level temporal information enhanced GCN (STIG) subnetwork, which is designed to capture the large-scale discriminative features of land surfaces, in which temporal information enhanced GCN (TIGCN) not only extracts spatial-spectral features of images, but also captures temporal correlations within satellite time series, thereby reducing the pseudochanges caused by temporal dynamics; second, a pixel-level 3D-CNN (P3DC) subnetwork, which is proposed to focus on subtle features by extracting temporal-spatial-spectral information to correct detection errors caused by superpixel segmentation; third, a gated knowledge dynamic selection module, which is designed to dynamically select the optimal features in the spectral dimension from the superpixel-level, pixel-level, and their combined features, thereby reducing the impact of redundant information on the detection performance. TSCD not only extracts large-scale and subtle features simultaneously, but also captures the temporal correlations among images. Comparative experimental results with eight state-of-the-art CD methods on nine publicly available satellite time series scenes demonstrate the effectiveness of the TSCD method.

Keywords

• Change detection
• dynamic selection
• graph convolutional network (GCN)
• satellite time series
• temporal information