Siamese Attentive Convolutional Network for Effective Remote Sensing Image Change Detection

Recently, deep learning-based architectures have delivered their great effectiveness in change detection (CD) of remote sensing (RS) images. However, most previous methods are lacking of extracting the multi-scale and long-range information due to the limitation of convolutional operations. To tackle these problems, in this work we propose a fresh framework named Siamese Attentive Convolutional Network (SACNet) for effective RS image change detection. Technically, our proposed framework first utilizes a Siamese Fully Convolutional Network (SFCN) as the encoding stage for multi-scale feature extraction from bitemporal RS images. To improve the representation ability, we propose a Fusing Attention Module (FAM) to highlight and aggregate multi-level attentive features from the SFCN. Besides, to capture more contextual information, we also sequentially introduce an Atrous Spatial Pyramid Pooling (ASPP) module and a Squeeze-and-Excitation (SE) module to enlarge the receptive field and select the channel-refined features. As for the decoding stage, we introduce a Difference Attention Module (DAM) to select more discriminative features for the CD map prediction. The resulting structure can fully utilize the difference information to enforce the detection of change regions. Finally, we apply deep supervision on the multi-level side-outputs to achieve a better model optimization. Extensive experiments demonstrate the effectiveness of our proposed method on three public remote sensing image CD benchmarks. We will later release the source code for model reproduction.