RS-NormGAN: Enhancing change detection of multi-temporal optical remote sensing images through effective radiometric normalization

Radiometric normalization (RN), also known as relative radiometric correction, is usually utilized for multi-temporal optical remote sensing image pairs. It is crucial to applications including change detection (CD) and other time-series analyses. Nevertheless, the variations across multi-temporal remote sensing image pairs are complex, containing true changes of landcover and fake changes caused by observation conditions, which poses significant difficulties for CD and other applications. For CD, the goal of RN is to well eliminate the unwanted fake changes. However, neither traditional methods nor current deep learning methods offer satisfactory solution for multi-temporal remote sensing images RN when dealing with such complicated circumstances. Towards this end, a novel pseudo invariant feature (PIF)-inspired weakly supervised generative adversarial network (GAN) for remote sensing images RN, named RS-NormGAN, is proposed to improve CD efficiently. Motivated by PIF, a sub-generator structure with different constraints is introduced to adequately deal with variant and invariant features, respectively. Besides, a global–local attention mechanism is proposed to further refine the performance by compensating spatial distortion and alleviating over-normalization and under-normalization. To verify the effectiveness of RS-NormGAN, massive experiments for CD and semantic CD across diverse scenarios have been conducted on Google Earth Bi-temporal Dataset and a constructed benchmark Sentinel-2 Hefei Change Detection Dataset. Compared with state-of-the-art methods, the proposed RS-NormGAN is very competitive, even if a simple CD network is utilized. The data and code will be available at https://gitbub.com/lixinghua5540/RS-NormGAN.