ADS-Net: An Attention-Based deeply supervised network for remote sensing image change detection

Abstract Change detection technology is an important key to analyze remote sensing data and is of great significance for accurate comprehension of the earth's surface changes. With the continuous development and progress of deep learning technology, fully convolutional neural networks are applied gradually in remote sensing change detection tasks. The present methods mainly encounter the problems of simple network structure, poor detection of small change areas, and poor robustness since they cannot completely obtain the relationships and differences between the features of bi-temporal images. To solve such problems, we propose an attention mechanism-based deep supervision network (ADS-Net) for the change detection of bi-temporal remote sensing images. First, an encoding–decoding full convolutional network is designed with a dual-stream structure. Various level features of bi-temporal images are extracted in the encoding stage, then in the decoding stage, feature maps of different levels are inserted into a deep supervision network with different branches to reconstruct the change map. Ultimately, to obtain the final change detection map, the prediction results of each branch in the deep supervision network are fused with various weights. To highlight the characteristics of change, we propose an adaptive attention mechanism combining spatial and channel features to capture the relationship of different scale changes and achieve more accurate change detection. ADS-Net has been tested on the LEVIR-CD and SVCD datasets of challenging remote sensing image change detection. The results of quantitative analysis and qualitative comparison indicate that the ADS-Net method comprises better effectiveness and robustness compared to the other state-of-the-art change detection methods.