Forecasting satellite-retrieved land surface temperature from reanalysis with multi-modal deep learning

Timely estimates of land surface temperature (LST) are critical in weather and climate prediction. Examples include assessing effects of extreme heat and drought on the biosphere and modelling transport processes in the atmospheric boundary layer. Yet, forecasting the spatiotemporal variability of LST remains challenging because the surface skin responds to forcing instantaneously and is controlled by multi-scale thermodynamic processes. Existing work on surface temperature forecasting largely follows two distinct paradigms: A) AI-driven and numerical weather prediction where large-scale skin temperature is simulated from Earth system models or their emulators, and B) geoscientific remote sensing where satellite-retrieved LST is extrapolated in time or space on small spatial scales, including site-scale experiments, often using statistical autoregression. While the goal of A) is to provide global estimates and atmospheric boundary conditions on coarse resolution with reduced complexity of subgrid processes, the goal of B) is often to obtain better forecasts over limited areas or local stations for downstream applications but these approaches rarely incorporate synoptic-scale meteorological context.

Large-scale approaches of medium-complexity to surface temperature forecasting that bridge these two ends and account for synoptic-scale surface meteorology while being sensitive to local land conditions remain underexplored. One reason for this is that modeling the tight coupling of spatial heterogeneity to multi-scale surface energy balance processes requires incorporation of multiple data sources at different spatiotemporal resolution. We cross these two paradigms and develop an observation-guided system that produces short-term forecasts of LST from reanalysed, coarse resolution surface meteorology and ancillary geostationary-resolution land surface properties. This system will cover the diurnal cycle and spatially larges scales at geostationary-resolution. We leverage the possibilities of multi-modal supervised learning and incorporate both reanalysis and observational data, explore memoryless and autoregressive approaches and outline opportunities to include high-resolution observations. Our approach is a first step towards effectively downscaling forecasts from the WeatherGenerator foundation model to high resolution surface conditions.