Estimating Rainfall with Multi-Resource Data over East Asia Based on Machine Learning

The lack of accurate estimation intense precipitation is a universal limitation in retrieval. Therefore, new rainfall retrieval technique based on the Random Forest (RF) algorithm presented using Advanced Himawari Imager-8 (Himawari-8/AHI) infrared spectrum data and NCEP operational Global Forecast System (GFS) forecast information. And gauge-calibrated estimates from Precipitation Measurement (GPM) product served as ground truth to train model. two-step RF classification model was established for (1) rain area delineation (2) grades’ improve accuracy moderate heavy rain. In view imbalance categories’ distribution datasets, resampling including Under-sampling Synthetic Minority Over-sampling Technique (SMOTE) implemented throughout whole training process fully learn characteristics among samples. Among features used, contributions meteorological variables trained models were generally greater than those information; particular, contribution precipitable water largest, indicating sufficient necessity vapor conditions forecasting. simulation results by compared with GPM pixel-by-pixel. To prove universality model, we used independent validation sets which are not two testing different periods set. addition, validated against gauge GFS rainfall. Consequently, identified areas Probability Of Detection (POD) around 0.77 False-Alarm Ratio (FAR) 0.23 validation, well POD 0.60–0.70 FAR 0.30 testing. estimate grades, value 0.70 0.60 despite certain overestimation. summary, performance test indicated great adaptability superiority East Asia. extent, our study provides meaningful range division powerful guidance quantitative estimation.