A stepwise unmixing model to address the scale gap issue present in downscaling of geostationary meteorological satellite surface temperature images

Land surface temperature (LST) is a critical parameter that drives the response of a variety of ecosystems to environmental and climatic changes. The geostationary satellite brings unique opportunities to monitor the LST at a hemispheric scale with temporal resolutions of up to 5 min. However, the ultra-coarse spatial resolutions ranging from 2 km to 5 km limit its application at local spatial scales. Downscaling the geostationary satellite LST image with the high-resolution low-Earth-orbit satellite images is a cost-effective way to circumvent this dilemma. Yet, the big gap between the observation scales of these satellite data poses a challenge for accurate downscaling. To address this problem, we proposed a stepwise temperature unmixing (TUM) model called ‘UnmixGO’, which downscales hourly LST images of Geostationary Operational Environmental Satellites (GOES-R) from 2 km to 100 m resolution. The spatially adaptive endmembers and the constrained solution space of the TUM model keep the errors in downscaled LSTs from being over-amplified in the stepwise data treatment. We validated the algorithm in six experimental areas and at five flux tower sites across the contiguous United States, revealing that UnmixGO outperformed conventional methods in accuracy by 0.49 K and 1.11 K on average for downscaling the simulated and real GOES-R LST images, respectively. Furthermore, the technical framework employed by UnmixGO is compatible with multi-source satellite images, enhancing the added value of our study in a future where various remote sensing data is increasingly accessible.