Enhancing subseasonal forecasting skill with land observations and physics-informed deep learning

Subseasonal forecasts, which predict weather patterns from weekly up to seasonal timescales, are crucial to minimize the adverse impacts of extreme weather events, such as heatwaves and droughts, on ecosystems and society. However, forecast skill at subseasonal lead times remains limited, as the chaotic nature of the atmosphere reduces the usefulness of the information contained in atmospheric initial conditions for increasing lead times. In contrast, land surface states, including soil moisture and vegetation anomalies, evolve more slowly and retain memory over weeks, allowing them to persist across subseasonal timescales and making them a potentially important source of predictability. Despite this, most operational weather forecast models represent only the mean seasonal cycle of land conditions, because accurately incorporating land surface anomalies remains challenging and can degrade model performance.

In order to address this situation, we develop a prototype of a hybrid weather prediction model to forecast near-surface temperature and surface soil moisture and related extremes. The model leverages the flexibility of deep learning to build on (i) satellite-based land surface observations, (ii) short-range forecasts from the Integrated Forecasting System to inform the model with physically consistent atmospheric evolution, and (iii) previous meteorological conditions sourced from reanalysis data. First results suggest that land surface anomalies exert a stronger influence during extreme conditions, when land memory persists, whereas under average conditions their influence is more evenly shared with atmospheric anomalies. Our study provides a benchmark for integrating land surface information into hybrid forecasting systems and highlights pathways to improve subseasonal prediction and early warning systems.