Bias Correction of Satellite Precipitation Products Near the Data-Scarce Region of Al-Ahsa, Saudi Arabia

Accurate estimation of precipitation is critical for water resource management under changing climate dynamics. This requirement is further intensified in data-scarce arid regions where ground-based observations are limited and sparsely distributed. Gridded satellite precipitation products offer consistent spatial and temporal coverage. However, they often suffer from systematic and random biases due to limitations in retrieval algorithms. This study aims to establish an effective bias correction algorithm in arid regions by evaluating the performance of three recommended machine learning models, namely, random forest, XGBoost, and ConvLSTM, over the area surrounding Al-Ahsa oasis in Saudi Arabia. CMORPH and IMERG datasets, along with climate variables (minimum temperature, relative humidity, and cloud fraction) from ERA-5, were compared with ground observations for model calibration and validation. Statistical metrics, including the correlation coefficient, root mean square error, mean absolute error, and relative bias, were used to assess the performance. The ConvLSTM model outperformed Random Forest and XGBoost, reducing the relative bias from 50.8% in raw CMORPH and 282.2% in raw IMERG to -1.71% and +1.15%, respectively. Graphical evaluation of daily average precipitation also supported the results indicated by statistical metrics. This study suggests that the one-dimensional ConvLSTM model was particularly effective in regions with limited topographical variation, making it suitable for bias correction in arid areas like Al-Ahsa, where temporal variability is prominent. The corrected precipitation products developed through this research are an invaluable asset for sustainable water resource management, agricultural planning, and drought mitigation in similar arid regions.