ML-based diagnostics of spatial precipitation maximum using large-scale atmospheric predictors in midlatitudes

This study is devoted to the problem of statistical downscaling of summer precipitation using the example of the Moscow region. We propose an approach to downscaling the local maximum of daily precipitation amounts within the region using machine learning (ML) models based on predictors characterizing large-scale meteorological processes. The study is based on 33 years of daily maximum precipitation data collected from 27 weather stations in the Moscow region used for ML model training and evaluation. We propose a set of physically justified precipitation predictors from ERA5 reanalysis (averaged over the region) including basic atmospheric variables (temperature, humidity, etc.) at different vertical levels as well as more complex indices characterizing convective instability, wind shear, humidity and circulation indicators. We evaluate three different ML models and several configurations of the feature selection and processing. The gradient boosting model with the daily averaged, standardized predictor set, including the reanalysis precipitation, demonstrated the best quality, with RMSE of 8.47 mm and R ² of 0.6. The feature importance analysis revealed that the mean precipitation from reanalysis as well as several complex indices are the key influencing factors for precipitation maxima.