Enhancing satellite based precipitation estimates through robust merging frameworks and streamlined ensemble machine learning model development

Accurate estimation of daily precipitation is essential for effective water resource management and climate risk assessment. Satellite precipitation products (SPPs) offer valuable spatial coverage but remain limited by uncertainties, particularly in arid and semi-arid regions. To address these challenges, this study develops a robust merging framework that integrates four SPPs with auxiliary topographic and meteorological data using ensemble machine learning models (EMLMs). Within this framework, we introduce for the first time the Multiple Linear Regression–based Sine Cosine Algorithm (MLR-SCA), designed to improve merging performance relative to the widely used Bayesian Model Averaging (BMA). Daily precipitation observations from 80 synoptic stations across Iran (2014–2022) were employed for training and validation. Results demonstrate that the proposed MLR-SCA significantly outperforms BMA, increasing the correlation coefficient (CC) by 132%, reducing RMSE by 34% and MAE by 19%, and achieving substantial improvements in KGE (+ 1142%), POD (+ 40%), and CSI (+ 47%), while reducing FAR (–24%) and BIAS (–7%). Although merging slightly reduced categorical event-detection skill in some cases, the EMLM framework consistently produced more accurate, stable, and reliable precipitation estimates across diverse climatic zones. Compared with existing approaches, the proposed framework offers three main advantages: (1) stronger performance across arid, semi-arid, and semi-humid climates; (2) improved detection of extreme precipitation events, which are often underestimated by raw SPPs; and (3) greater robustness through the simultaneous integration of multiple SPPs and auxiliary datasets. These findings highlight the potential of the EMLM–MLR-SCA framework to support operational hydrology, water resource planning, and climate adaptation in data-scarce regions.