Precision in Precipitation: Bias Corrections and Machine Learning for Reliable Satellite Precipitation in The Levant

Reliable precipitation data from in-situ stations is often limited by inconsistent quality, resolution, and spatial coverage. This is particularly true in regions like the West Bank, where ground-based observations are scarce. This hampers hydrological and environmental studies where accurate precipitation estimates are vital.  Therefore, satellite-based rainfall products are an appealing alternative due to their broad spatial and consistent temporal coverage. However, the accuracy of these products in complex terrain is questionable due to sensor and retrieval errors, necessitating adjustment to improve their reliability. This study evaluates various adjustment methods for four satellite precipitation products (IMERG Final Run, PDIR-Now, CCS-CDR, and CMORPH) across the study area of Historical Palestine (West Bank and Israel). Daily satellite precipitation estimates were compared to observations from 316 in-situ stations (256 in Israel and 58 in the Palestinian territories). Adjustment methods included traditional bias correction techniques (Linear Scaling, Daily Translation, and Annual Sums), more advanced approaches (Empirical Quantile Mapping, Robust Quantile Mapping, Gaussian Distribution Mapping, and Local Intensity Scaling), and machine learning models (Random Forest and Artificial Neural Networks). Results show that, among the non-machine learning approaches, Daily Translation (DT) achieved the greatest improvement in accuracy followed by Power Bias adjustment. DT applied to IMERG resulted in an improvement of 24% and 17% in R² and Mean Absolute Error (MAE) respectively. All machine learning approaches outperformed non-machine learning methods, with a two-step Random Forest (RF2) method delivering the best results. RF2, which leverages data from multiple satellites, had a 109% improvement in R² and a 54% improvement in MAE. Additionally, the global RFG model showcased excellent results in producing a unified model that can be generalized for the entirety of the study area. The findings are globally applicable and evaluate multiple adjustment methods which opens the opportunity for easily accessible remotely sensed precipitation products to be used in many hydrological applications.