Dynamic Context Aggregation Based Change Detection Method for Building Remote Sensing Images

Aiming at the problems that the standard convolutional operation is difficult to effectively capture global features due to the local receptive field limitation and the selfattention mechanism does not fully consider the interrelationships of neighboring features, a change detection framework DCA-NET is proposed, which includes a twin encoder, a decoder, a fusion module, and a Dynamic Contextual Attention (DCA) module. The DCA module effectively fuses static and dynamic contextual representations. It then leverages the static contextual information to guide the learning of the dynamic attention matrix. A new coordinate generation algorithm is introduced to define the initial position of the convolutional kernel of arbitrary size, so that the shape and size of the convolutional kernel can be flexibly adjusted, so that the model can better capture the global structure of large-scale and the detailed features of small scale, thus improving the perceptual ability of the detection model. Experiments were conducted on three public CD (Change Detection) datasets; LEVIR-CD, WHU-CD, and CDD-CD. 90.04%, 88.12% and 93.87% F1 scores were obtained respectively. The experimental results show that the DCA-NET model achieves significant performance improvement on the change detection task with strong stability and generalization capabilities.