MSWEP V3: Enhancing Global Precipitation Estimates with Machine Learning

We introduce Version 3 (V3) of the gridded near real-time Multi-Source Weighted-Ensemble Precipitation (MSWEP) product — the first fully global, machine learning-based precipitation (P) product, developed to address the growing demand for accurate precipitation data amid escalating climate challenges. MSWEP V3 provides hourly 0.1° resolution data from 1979 to the present, updated continuously with a latency of less than two hours. The development involves a two-stage process: first, baseline P fields are generated using machine learning model stacks that integrate satellite and (re)analysis P and air temperature products alongside static P-related variables, trained with hourly and daily observations from nearly 18,000 global gauges. Second, these fields are corrected using available daily gauge observations, accounting for gauge reporting times. To assess MSWEP V3's performance, we conducted an extensive evaluation of 19 gridded P products, using independent observations from almost 18,000 gauges excluded from training. MSWEP V3 (prior to gauge corrections) achieved a mean daily Kling-Gupta Efficiency (KGE) value of 0.69, outperforming all 18 other products evaluated. For comparison, other non-gauge-corrected products such as CHIRP, ERA5, GSMaP V8, and IMERG-L V7 achieved mean KGE values of 0.31, 0.61, 0.38, and 0.46, respectively. MSWEP V3 consistently ranked first or second across multiple metrics, including correlation, overall bias, peak bias, wet days bias, and the critical success index. Notably, MSWEP V3 (without gauge corrections) also outperformed several products that directly incorporate gauge observations, such as CHIRPS, CPC Unified, and IMERG-F V7, which achieved mean KGE values of 0.36, 0.54, and 0.62, respectively. Set for release in early 2025, we anticipate that MSWEP V3 will support climate research, water resource assessments, flood management, and numerous other applications.