Thick Cloud Removal in Multitemporal Remote Sensing Images Using a Coarse-to-Fine Framework
Yue Zi,Xuedong Song,Fengying Xie,Zhiguo Jiang
DOI: https://doi.org/10.1109/lgrs.2024.3378691
IF: 5.343
2024-03-27
IEEE Geoscience and Remote Sensing Letters
Abstract:Remote sensing (RS) images are widely used for Earth observation. However, cloud contamination greatly degrades the quality of RS images and limits their applications. In this letter, we propose a coarse-to-fine thick cloud removal method for a single pair of multitemporal RS images. First, we perform a global color transformation on a cloud-free reference image using linear regression coefficients between the pixels in the cloudy target image and the reference image in the same cloud-free regions and obtain a coarse result. Then, a convolutional neural network (CNN) based on internal constraint is used to refine the coarse result, which does not require any construction of additional external training dataset in advance. We further design a multiscale feature extraction and fusion module and an auxiliary loss involving cloud regions to improve the performance of the CNN. Finally, Poisson image fusion is used to generate a seamless cloud-free result. On a simulated test set containing 500 pairs of multitemporal RS images, the proposed method achieves satisfactory results with 25.1277 dB in peak signal-to-noise ratio (PSNR), 0.9077 in structural similarity (SSIM), and 0.9342 in correlation coefficient (CC). Qualitative and quantitative comparisons of our proposed against several state-of-the-art methods on the simulated and real cloudy images demonstrate the superiority of the proposed method.
imaging science & photographic technology,remote sensing,engineering, electrical & electronic,geochemistry & geophysics