Enhancing GRACE-derived terrestrial water storage estimates through bias-corrected downscaling using machine learning and artificial bee colony optimization

Accurate monitoring of spatiotemporal variations of Terrestrial Water Storage (TWS) is essential for sustainable water resources management in arid and semi-arid regions. The Gravity Recovery and Climate Experiment (GRACE) and GRACE-Follow On (GRACE-FO) satellite missions serve as valuable tools for estimating changes in TWS; however, their coarse spatial resolution poses challenges for applications at local and regional scales. In this study, downscaling of the GRACE data for the Yazd province, Iran, was conducted using three Machine Learning (ML) models, including Extreme Gradient Boosting (XGBoost) and two Deep Neural Networks, i.e., DNN1 and DNN2, while the latter has more hidden layers than the former. To enhance accuracy and reduce the systematic bias in the model outputs, the Artificial Bee Colony (ABC) metaheuristic algorithm and a hybrid version combining ABC and Quantile Mapping were employed as the bias correction framework. The input data for the models were derived from high-resolution datasets, including CHIRPS for precipitation and TerraClimate for climatic variables listed in Table 1. The performance evaluation of the models, based on statistical indices, indicated that the XGBoost model achieved the highest accuracy in reproducing TWS anomalies (R2 = 0.829) and the lowest error among the ML models. The implementation of the ABC algorithm resulted in a 12.45% reduction in root mean square errors (RMSE) and improved the consistency between model outputs and observations from piezometer wells. Temporal and spatial analyses demonstrate a continuous decline in equivalent water thickness over the past two decades, confirming a severe depletion of groundwater resources in Yazd. The findings of this study demonstrate that the integration of ML models with ABC can enhance model performance by reducing prediction errors and serve as an effective approach for generating higher-resolution GRACE-derived TWS estimates at the local scale, thereby supporting more informed water resources management decisions in arid regions. Improved downscaling of GRACE-derived using XGBoost, deep learning and multi-layer perceptron models to produce higher-resolution Terrestrial Water Storage maps in Yazd Province. Utilized high-resolution datasets, including CHIRPS and TerraClimate, as main inputs for the downscaling models. Reduced model bias and enhanced agreement with piezometer observations wells using the Artificial Bee Colony algorithm.