ChangeRD: A registration-integrated change detection framework for unaligned remote sensing images

Abstract

Change Detection (CD) is important for natural disaster assessment, urban construction management, ecological monitoring, etc. Nevertheless, the CD models based on the pixel-level classification are highly dependent on the registration accuracy of bi-temporal images. Besides, differences in factors such as imaging sensors and season often result in pseudo-changes in CD maps. To tackle these challenges, we establish a registration-integrated change detection framework called ChangeRD, which can explore spatial transformation relationships between pairs of unaligned images. Specifically, ChangeRD is designed as a multitask network that supervises the learning of the perspective transformation matrix and difference regions between images. The proposed Adaptive Perspective Transformation (APT) module is utilized to enhance spatial consistency of features from different levels of the Siamese network. Furthermore, an Attention-guided Central Difference Convolution (AgCDC) module is proposed to mine the deep differences in bi-temporal features, significantly reducing the pseudo-change noise caused by illumination variations. Extensive experiments on unaligned bi-temporal images have demonstrated that ChangeRD outperforms other SOTA CD methods in terms of qualitative and quantitative evaluation. The code for this work will be available on GitHub.