Multiple Attention Siamese Network for High-Resolution Image Change Detection

Abstract

Change detection (CD) remains an important issue in remote sensing applications, especially for high-resolution images, but it has yet to be fully resolved. In this study, we propose a novel Siamese network model, i.e., the multiple attention Siamese network (MASNet), for high-resolution image change detection (HRCD). The selective kernel convolution (SKConv) is first embedded into the encoders of the Siamese network to improve its feature extraction ability. The attention feature fusion module (AFFM) is then proposed and utilized to reform the network’s decoder, which realizes the fusion and selection of image features extracted from different scales and times. Experiments on three benchmark CD datasets, namely, the season-varying change detection dataset (SCDDS), Wuhan University (WHU) building change detection dataset (WBDS), and Google dataset (GDS), are conducted to validate the proposed model. The results show that the proposed attention module (the AFFM) is superior to multiple classic attention modules, including the squeeze-and-excitation block (SE), spatial SE block (sSE), spatial and channel SE block (scSE), convolutional block attention module (CBAM), and efficient channel attention module (ECA), in terms of the fusing of CD features. Furthermore, MASNet achieves a higher overall accuracy (OA), mean intersection over union (mIoU), F1-score (F1), and kappa coefficient (kappa) than ten state-of-the-art HRCD methods, such as the deeply supervised image fusion network (DSIFN), spatial–temporal attention neural network (STANet), and dual attentive fully convolutional Siamese network (DASNet) while maintaining low missed alarms and false alarms and having a fair method efficiency.