Integrated Methodology for Atmospheric Correction and Cloud Removal of Multispectral Remote Sensing Images Using Guided Diffusion Model

The atmospheric correction and cloud removal are core tasks of the atmospheric radiative transfer model in satellite remote sensing radiometric calibration. The accurate atmospheric correction and cloud removal of satellite remote sensing images are critical preliminary steps for effective satellite remote sensing applications. Current methods for atmospheric correction and cloud removal in satellite remote sensing imagery are mainly divided into traditional algorithms and deep learning methods. Traditional remote sensing algorithms typically involve complex atmospheric radiative transfer models, intricate meteorological parameters, synchronized ground measurements, and corresponding geographic data. There are significant difficulties and limitations in achieving detailed descriptions of ground and atmospheric models. Deep learning methods typically employ computer vision techniques. These methods cannot effectively utilize atmospheric parameters and multispectral image (MSI) features. Therefore, they cannot significantly improve the data quality of spectral satellite remote sensing images. This study integrates multitemporal and multispectral remote sensing images, atmospheric data (such as aerosol optical thickness and water vapor), and remote sensing data like surface reflectance (SR). A novel unified model is proposed to handle atmospheric correction and cloud removal for multispectral satellite remote sensing images. The unique geographical environment, and diverse and complex atmospheric conditions of the Guilin region make its satellite remote sensing data rich in research characteristics for atmospheric correction and cloud removal tasks. It can provide a typical and representative experimental scenario for this study. This study first utilized Sentinel-2 top-of-atmosphere (TOA) and SR data from the Guilin region to create four datasets: true color and multispectral fusion datasets with atmospheric correction incorporating aerosol parameters, single and multitemporal cloud removal datasets. Then, we developed a conditional guided diffusion model. This model can capable for both multispectral atmospheric correction and multitemporal cloud removal tasks by redesigning the conditional control and encoding sections. Our method achieved satisfactory results in both true color and multispectral atmospheric correction tasks. For atmospheric correction tasks on true color images (TCIs), our method achieved MAE, SSIM, and peak signal-to-noise ratio (PSNR) metrics of 0.0125, 0.9445, and 35.5221 dB, respectively. For the atmospheric correction task in MSIs, the method achieved MAE, SSIM, and PSNR metrics of 0.0174, 0.9499, and 36.3547 dB, respectively. For multitemporal true color cloud removal tasks, the method achieved MAE, SSIM, and PSNR metrics of 0.0224, 0.9556, and 36.2122 dB, respectively.