RACDNet: Resolution- and Alignment-Aware Change Detection Network for Optical Remote Sensing Imagery

Change detection (CD) methods work on the basis of co-registered multi-temporal images with equivalent resolutions. Due to limitation sensor imaging conditions and revisit period, it is difficult acquire desired images, especially in emergency situations. In addition, accurate co-registration largely limited by vast object changes matching algorithms. To this end, a resolution- alignment-aware change network (RACDNet) proposed for multi-resolution optical remote-sensing imagery CD. first stage, generate high-quality bi-temporal light-weighted super-resolution fully considering construction difficulty different regions, which facilitates detailed information recovery. Adversarial loss perceptual are further adopted improve visual quality. second deformable convolution units embedded novel Siamese–UNet architecture deep features alignment; thus, robust difference can be generated extraction. We use an atrous module enlarge receptive field, attention bridge semantic gap between encoder decoder. verify effectiveness our RACDNet, dataset (MRCDD) created using Google Earth. The quantitative qualitative experimental results demonstrate that RACDNet capable enhancing details reconstructed significantly, performance CD surpasses other state-of-the-art large margin.