Long-term precipitation prediction in different climate divisions of California using remotely sensed data and machine learning

ABSTRACTThis study presented a novel paradigm for forecasting 12-step-ahead monthly precipitation at 126 California gauge stations. First, the satellite-based precipitation time series from Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS), TerraClimate, ECMWF Reanalysis V5 (ERA5), and Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Climate Data Record (PERSIANN-CDR) products were bias-corrected using historical precipitation data. Four methods were tested, and quantile mapping (QM) was the best. After pre-processing data, 19 machine-learning models were developed. random forest, Extreme Gradient Boosting (XGBoost), extreme gradient boosting, support vector machine, multi-layer perceptron, and K-nearest-neighbours were chosen as the best models based on Complex Proportional Assessment (COPRAS) measurement. After hyperparameter adjustment, the Bayesian back-propagation regularization algorithm fused the results. The superior models’ predictions were considered inputs, and the target’s initial step was labeled. The next 11 steps at each station followed this approach, and the fusion models accurately predicted all steps. The 12th step’s average Nash-Sutcliffe efficiency (NSE), mean square error (MSE), coefficient of determination (R2), correlation coefficient (R) were 0.937, 52.136, 0.880, and 0.869, respectively, demonstrating the framework’s effectiveness at high forecasting horizons to help policymakers manage water resources.KEYWORDS: bias correctionhyperparameterslong-term precipitation predictionmachine learning (ML)quantile mapping (QM)satellite-based precipitation Editor A Castellarin; Associate Editor F-J. ChangEditor A Castellarin; Associate Editor F-J. ChangDisclosure statementNo potential conflict of interest was reported by the authors.Supplementary materialSupplemental data for this article can be accessed online at https://doi.org/10.1080/02626667.2023.2248112.