Denoising Diffusion Probabilistic Models for Cloud Removal and Land Surface Temperature Retrieval From a Single Sample

Land Surface Temperature (LST) monitoring is a key factor in demonstrating global climate change. Considering that cloud cover often obscures accurate remotely sensed LST data, it is important to develop approaches to reconstruct occluded areas. This paper presents an inpainting technique, designated RePaint, which utilizes Denoising Diffusion Probabilistic Models (DDPM) for the removal of clouds and the retrieval of LST data from Landsat-8 remote sensing imagery. The DDPM was trained with cloud-free LST image patches, and binary masks were formed by the segmentation process using the Ukis Cloud Shadow Mask to condition the region to be reconstructed. To simulate cloud occlusion in the test data, three cloud cover scenarios with 15%, 22% and 50% missing rate data were investigated. The performance of the approach was evaluated through a combination of quantitative and qualitative assessments. The quantitative analysis showed that the model exhibited consistent performance during the reconstruction process, achieving a mean squared error 0.0001, a peak signal to noise ratio 41.2888, and a structural similarity index 0.9826. In order to provide a more comprehensive visual representation of the results, boxplots and the Gaussian distribution of the temperature error in the LST reconstruction were included. Images of the absolute error of the reconstruction and scatterplots were used for qualitative analysis of randomly selected LST patches. The results demonstrate that the characteristics of the land surface have been preserved, exhibiting a minimum incidence of temperature reconstruction error and spatial coherence across all scenarios.