A Deep Learning-based Land-Atmosphere Coupled Model for Heatwave Prediction

Abstract Extreme heat events are intensifying under climate change, yet their prediction remains limited by the inadequate representation of land–atmosphere (L–A) interactions. Most deep learning–based weather models have focused solely on atmospheric variables, overlooking the role of land surface conditions in driving heat extremes. In this study, we present an L–A coupled prediction framework to enhance heatwave forecasting, focusing on Northern Hemisphere summer conditions. The model integrates soil moisture and temperature across multiple vertical layers into the atmospheric forecasting process and is trained to reflect their dynamic influence. We found that training the model with a multi-step loss function significantly improved its ability to capture L–A interactions on a sub-seasonal time scale by sustaining their strength and structure across longer lead times. Using multi-step loss, the L–A coupled model achieved a 5.9–11.2% improvement in heatwave forecast accuracy relative to the atmosphere-only model across 1–7 day lead times, as measured by root mean squared error, whereas the improvement was only 1.9–4.3% with single-step loss. The coupled model’s forecast skill gain is strongest at short leads (~ 3 day) when both SM and circulation predictability were high, and sustained at longer leads (up to 7 day) mainly by L–A coupling through SM predictability. Case studies of recent WesternEuropean and East Asian heatwaves further demonstrated its ability to capture land surface drying and associated temperature extremes. These findings underscore the importance of incorporating L–A coupling with multi-step optimization in advancing data-driven heatwave forecasting systems.