思维链赋能多模态遥感影像云智能识别与修复OA

To address the limitation of existing methods that isolate cloud detection and thin/thick cloud restoration in remote sensing images and lack a unified reasoning framework,this paper proposed a chain-of-thought(CoT)-based intelligent cloud processing model for multimodal remote sensing images.The model decomposed the task into three stages:"cloud detection,thin cloud removal,and thick cloud inpainting".First,a semantic segmentation network was adopted to generate precise cloud masks and classify cloud thickness;then,a modified denoising diffusion probabilistic model(DDPM)was employed to restore information in thin cloud regions;finally,thick cloud inpainting was completed by fusing synthetic aper-ture radar(SAR)and optical data through a flow matching(FM)-based generative model.Quantitative and qualitative experiments on Gaofen-2 imagery of the Hunan region demonstrate that for thin cloud removal,the proposed method achieves a peak signal-to-noise ratio(PSNR)of 30.3 dB,a structural similarity index(SSIM)of 0.928,and a root mean square error(RMSE)of 0.029;for thick cloud inpainting,it achieves a PSNR of 27.9 dB,an SSIM of 0.894,and a Fréchet inception distance(FID)as low as 25.4,all of which outperform mainstream methods.The results validate the effec-tiveness and advancement of the CoT framework,providing an efficient and coherent technical solution for cloud processing in complex remote sensing images.