Land surface information from satellites boost near-surface temperature forecast skill

Abstract Earth observations from satellites are the primary information source for Numerical Weather Prediction models. While some land surface variables, such as surface soil moisture, are assimilated to improve initial land conditions, the use of additional satellite-derived land surface and vegetation products remains limited partly due to systematic model biases. Here we examine the potential of satellite-derived land surface temperature and vegetation indicators to enhance near-surface temperature forecast skill. We build deep learning surrogate models for Numerical Weather Prediction using Long Short-Term Memory networks. Results show that including these satellite datasets improves temperature forecast skill globally across lead times from 1 to 12 days, with the largest improvements at 4-day lead time. Satellite-based predictors are the most relevant variables in about 60% of global grid cells. Among them, sun-induced fluorescence is the most important predictor, reflecting vegetation photosynthetic activity and its influence on surface energy partitioning and near-surface temperature.