Change Detection of High-resolution Remote Sensing Images Through Adaptive Focal Modulation on Hierarchical Feature Maps

One of the major challenges in change detection (CD) high-resolution remote sensing images is high requirement for computational resources. Besides, to get best result, it must spot only important changes while omitting unimportant ones, which requires learning complex interactions between multi-scale objects on images. Despite Convolution Neural Network (CNN) efficiently extracting features from such images, has a limited receptive field resulting sub-optimal representation. On other hand, Vision Transformer (ViT) can capture long-range dependencies. Still, suffers quadratic complexity concerning number image patches, especially Furthermore, both approaches not model among essential understand natural fully. We propose FocalCD, CD method based recently proposed focal modulation architecture capable short and long dependencies solve this problem. It attention-free does suffer complexity. Also, supports interaction by adaptively selecting discriminator regions levels. Besides efficient yet powerful encoder, FocalCD an effective feature fusion pyramidal decoder network. achieves strong empirical results various datasets, including CDD, LEVIR-CD, WHU-CD. reaches F1 scores 0.9851, 0.952, 0.9616 datasets WHU-CD outperforming state-of-the-art methods having comparable or even lower computation