Work

Machine Learning Estimation of Central Java Monthly Rainfall from Bias Corrected NASA POWER

Year: 2026
Type: article
Abstract: Monthly rainfall is the working time step for water availability assessment, reservoir yield analysis, and irrigation planning; however, rain gauge records in Indonesia are sparse and often interrupted. This study compares three methods for estimating monthly rainfall in Central Java: multiple linear regression (MLR), eXtreme Gradient Boosting (XGBoost), and random forest (RF). Bias-corrected NASA POWER rainfall serves as the target, and NASA POWER air temperature, relative humidity, and wind speed serve as predictors; therefore, no rainfall information enters the models. Raw daily NASA POWER rainfall was corrected bin by bin against the empirical gauge probability-duration curve at nine stations. The correction raised the mean monthly gauge-satellite correlation from 0.762 to 0.810, with the largest gain at Maritim Tegal, from 0.429 to 0.607. Verification against gauge observations over 355 station-months (247 for MLR) gave RMSE values of 125.3, 87.1, and 79.8 mm/month and Nash-Sutcliffe efficiencies of 0.42, 0.74, and 0.78 for MLR, XGBoost, and RF, respectively. RF gave the lowest RMSE at six of the nine stations. On objective-function criteria previously used to rank global rainfall products in an Indonesian watershed, both ensembles earned the maximum score of 16, against 13 for MLR, and their four-class BMKG intensity hit rates were 75.5% and 75.2%, compared with 53.8% for MLR. Therefore, RF is recommended for rainfall totals that feed water balance or hydrological models, XGBoost for intensity-class outputs, and MLR as an interpretable benchmark,
Language: English
Open Access status: diamond
OpenAlex ID: W7233601952
FWCI:
Cites:
Cited by:
Related to: