Hourly 0.1° Gridded Near Real-Time Precipitation (1979–Present) via Machine Learning Fusion of Satellite, Model, and Gauge Data

We introduce a new version of the gridded near real-time Multi-Source Weighted-Ensemble Precipitation (MSWEP) product, developed to address the urgent need for accurate precipitation (P) data in the face of escalating climate change challenges. The product has an hourly 0.1° resolution spanning 1979 to the present, and is continuously updated, with a latency of approximately one hour. The development process involves two stages. Firstly, baseline P fields are generated from multiple satellite and (re)analysis P products, along with several static P-related variables, using random forest models trained on 3-hourly and daily P observations from gauges across the globe (n=17,322). Subsequently, these baseline P fields are locally corrected using available daily P observations, employing a procedure that accounts for the reporting times of gauges. To assess the accuracy of the product, we conducted the most comprehensive global evaluation of P products to date, using daily observations from independent P gauges as a reference (n=15,184). The new P product (prior to gauge corrections) outperformed all 18 other evaluated products, attaining a mean daily Kling-Gupta Efficiency (KGE) value of 0.65. In contrast, widely used products such as CHIRP, ERA5, GSMaP, and IMERG achieved mean KGE values of 0.31, 0.57, 0.37, and 0.40, respectively. Furthermore, our P product consistently ranked first or second across various metrics, including correlation, overall bias, peak bias, wet days bias, and critical success index. Notably, the new product also outperformed several gauge-based products like CHIRPS and CPC Unified, which had mean KGE values of 0.37 and 0.54, respectively. Set for release in late 2024, we anticipate that the new product will be useful for climate research, water resource assessment, and flood management, among numerous other potential applications.