Dual-Attention Cross Fusion Context Network for Remote Sensing Change Detection

Abstract

Detecting changes in two remote sensing images of the same region but at different times is of great significance in applications, such as land management and urban planning, which also prompts the continuous development and progress of change detection (CD) technology. The current deep learning-based methods make full use of the excellent feature learning ability of deep convolution to show excellent detection performance. However, advances in remote sensing technology also mean that detected objects have higher resolution and more complex content, which is more challenging for CD techniques. Strengthening the model's ability to learn the context of the detected remote sensing image can effectively improve the model's ability to distinguish between changing features and nonchanging features, thereby achieving higher-precision detection results. In order to explore and utilize the contextual information of different levels of features as much as possible, we design a dual-attention cross fusion module in our method to realize the cross-learning of contextual information of different scales during the decoding process. It will be able to complementarily fuse feature content of different granularities. We also propose an Atrous Pyramid Difference Module (APDM) to efficiently capture the difference information of two refined features by exploiting receptive fields of different sizes. In addition, in order to further improve the context modeling ability of the model, we introduce a context transformer block (Cot). Different from other transformer-based self-attention methods, Cot dynamically guides the learning of the attention matrix by the contextual information of the input keys. Our method achieves F1-scores of 91.08%/91.93%/79.80% on the LEVIR-CD/WHU-CD/DSIFN-CD datasets, respectively. Extensive qualitative and quantitative experiments on these datasets validate the effectiveness of our method.