Advances in multi-source data fusion for precipitation estimation: remote sensing and machine learning perspectives

Abstract

During an era defined by unparalleled climate variability and intensifying water security challenges, precise precipitation estimation plays a pivotal role in establishing resilient water management systems and effective disaster mitigation strategies. This study presents an in-depth review of recent advances in precipitation estimation using multi-source data, with an emphasis on integrating Remote Sensing (RS) techniques with advanced pattern recognition methods. Utilizing the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework, we aggregate findings from peer-reviewed studies published since 2010 to evaluate the efficacy of satellite platforms—comprising geostationary, low-Earth orbit (LEO), and hyperspectral sensors—in acquiring precipitation data. Multi-source datasets such as the Climate Prediction Center Morphing Method (CMORPH), Tropical Rainfall Measuring Mission (TRMM), Global Precipitation Measurement (GPM), and Fengyun-4B have been merged with state-of-the-art Machine Learning (ML) algorithms, including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, ensemble approaches (i.e., techniques that combine predictions from multiple models to improve performance and robustness), and hybrid architectures (i.e., frameworks that integrate fundamentally different model types, such as machine learning algorithms with physics-based simulations), to effectively capture the spatial and temporal variability of precipitation. These challenges are mitigated by incorporating high-resolution predictors, wherein Digital Elevation Models (DEM) alleviate orographic biases, the Normalized Difference Vegetation Index (NDVI) reflects land-surface interactions, and soil moisture supplies essential data for water balance calibration, thus rectifying spatial heterogeneity and sensor inaccuracies. These predictors are utilized in conjunction with sophisticated calibration methods grounded in Bayesian optimization and Transfer Learning (TL). Comparative assessments demonstrate that hybrid and ensemble models surpass traditional techniques, especially in hilly and data-scarce areas. The paper examines uncertainty reduction through enhanced pre-processing, multi-source integration, and novel downscaling-calibration frameworks, while emphasizing prospective future avenues such as physics-based neural networks and federated learning methodologies.