Skip to content
Review Open access

Centenary evolution of rainfall datasets: a comprehensive review of gauge, radar satellite, and physics-guided machine learning-based products

Nov 2026 · Environmental Research Communications · Vol 8 · 0 citations · 215 references
Physics

Abstract

Accurate precipitation data are crucial for understanding hydroclimatic processes to support water resource management and disaster risk reduction. This study presents a systematic review of rainfall datasets from the early 20th century to the present, emphasizing the roles of artificial intelligence (AI) and machine learning (ML) in precipitation estimation. Unlike prior reviews that primarily catalog global precipitation datasets or examine individual ML applications, this synthesis jointly traces observing-system evolution and evaluates when ML adds demonstrable value relative to named conventional baselines. The evolution of datasets is categorized into four overlapping phases: (i) gauge-based observations, (ii) ground-based weather radar quantitative precipitation estimation (iii) satellite-era products, and (iv) modern multi-source and AI-enhanced datasets. Early gauge-based datasets, provide long-term climate records vital for historical analysis, though they are limited by sparse spatial coverage and interpolation uncertainties. Ground-based radar networks added repeated kilometer-scale mapping of storm structure, although reflectivity-to-rainfall conversion and gauge adjustment introduce their own errors. The satellite era, marked by missions such as Tropical Rainfall Measuring Mission and global precipitation measurement, has significantly improved precipitation monitoring by providing near-global coverage with higher spatial and temporal resolution. However, these datasets still face challenges related to systematic biases and the representation of extreme events and complex terrains. Recent advancements highlight the important but conditional role of AI/ML techniques, including neural networks, convolutional architectures, and ensemble learning methods. Within specific study domains and validation designs, these approaches have improved precipitation estimation through applications such as satellite retrieval, bias correction, gap filling, downscaling, and multi-source data fusion. This study comparisons shows meaningful but non-universal gains: GPROF-NN reduced global precipitation measurements microwave imager retrieval mean absolute error from 0.079 to 0.043 mm h−1; a cloud-informed correction reduced IMERG early RMSE by 53%–66% in two Chinese basins and a two-step fusion framework improved modified KGE by 15%–85% across six source products. These values cannot be pooled as a formal meta-analysis because the studies use different references, climates, periods, and validation designs. The evidence does not establish universal transfer or consistent improvement in extreme-event representation; priorities are cross-region and forward-in-time validation, calibrated uncertainty quantification, and robustness under non-stationary climate conditions.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.