A classification‒regression‒correction (CRC) strategy for precipitation merging with gauge-independent spatial autocorrelation predictors

Multi-source precipitation merging enhances spatiotemporal estimation accuracy, yet machine learning (ML)-based methods suffer from limitations in rain/no-rain identification and violate ML principles by utilizing gauge-based spatial autocorrelation predictors. This study proposes an innovative classification‒regression‒correction (CRC) precipitation merging framework and constructs gauge-independent spatial autocorrelation predictors. The framework employs four ML models (KNN, RF, AdaBoost, and XGBoost) as candidate models in the classification and regression steps, with the correction step integrating their outputs to distinguish rain/no-rain events. Using the proposed method, this study carries out precipitation merging in the lower Yangtze River Basin and the Lixiahe region using gauge observations, IMERG, GSMaP, and ERA5. The results indicate that: 1) RF exhibits the highest accuracy in rain/no-rain classification, achieving a monthly average critical success index (CSI) of 0.684; 2) the RF–RF combination demonstrates the best performance, markedly improving precipitation detection and estimation capabilities, with the monthly average CSI increasing from approximately 0.450 to 0.688 and the Kling-Gupta Efficiency (KGE’) improving from about 0.500 to 0.725 compared to the parent datasets; and 3) the CRC merging strategy effectively enhances the ability to distinguish rain/no-rain events, while the constructed spatial autocorrelation predictors both prevent data leakage and enhance continuous accuracy. This study demonstrates the potential of the CRC framework to advance multi-source precipitation merging and provide robust technical support for enhancing meteorological disaster preparedness.