A residual-based approach to downscale all-sky LST using synthetic clear-sky priors

作者：Panagiotis Sismanidis, Frank Goettsche, Cheolhee Yoo, Iphigenia Keramitsoglou, Benjamin Bechtel · 发表于：Science of Remote Sensing · 年份：2026 · DOI：10.1016/j.srs.2026.100468 · 研究领域：Meteorological Phenomena and Simulations、Solar Radiation and Photovoltaics、Constraint Satisfaction and Optimization

Over the last decade, substantial progress has been made in the development of seamless all-sky Land Surface Temperature (LST) products. However, many agricultural and environmental applications require spatial resolutions finer than the currently available ≥ 1 km. A practical solution to this limitation is to downscale existing all-sky datasets by extending methods originally developed for clear-sky data to also account for cloudy-sky conditions. In this study, we propose such an approach. Our method first generates a synthetic LST image representing average clear-sky conditions at the target spatial resolution, that then adjusts using downscaled residuals derived from the difference between the all-sky and synthetic LST. This approach offers two key advantages: it ensures physically realistic, fine-scale LST patterns by using the synthetic LST image as an initial guess, and amplifies the local weather effects in the all-sky LST, making them easier for the model to learn. To evaluate the method, we use half-hourly all-sky LST from the Spinning Enhanced Visible and Infrared Imager (SEVIRI), which we downscale from ∼ 5 km to ∼ 1 km and validate against independent in-situ and satellite LST. The results demonstrate improved performance relative to the original data across a range of land covers, topographies, and climate conditions, while also highlighting the potential of a single random forest (RF) model to predict the LST under both clear and cloudy conditions.