Moisture Source–Receptor Relationships for Post-Processing Medium-Range Precipitation Forecasts

Abstract Data-driven post-processing of medium-range precipitation forecasts typically relies on forecasted predictors. However, incorporating observed predictors remains challenging because of the long temporal gap between present observations and future precipitation. This study proposes a physically interpretable approach that leverages atmospheric moisture transport, where source–receptor relationships persist until medium-range lead times. Moisture sources associated with each event and lead time were identified using FLEXPART, a Lagrangian-based particle transport model. The methodology pre-processes evaporation-related observed auxiliary predictors (surface temperature and wind speed) in the input feature dataset, retaining only features corresponding to the identified moisture sources, before training the machine learning model. The method was applied to spatial-mean daily precipitation forecasts (1–14-day lead times) over the Kyushu region of Japan during summers of 2016–2024 using eXtreme Gradient Boosting algorithm. The methodology demonstrates the potential to improve 4–6-day lead time forecasts, including an average improvement of Nash-Sutcliffe efficiency (NSE) of 15.1% relative to raw forecasts, with NSE transitioning into the positive regime at mid-medium-range lead times. However, under operational conditions, the NSE improvement decreased to 5.6%. These findings highlight the importance of remote moisture sources in precipitation forecasting.