MSWEP V3: Machine Learning-Powered Global Precipitation Estimates at 0.1$^\circ$ Hourly Resolution (1979-Present)

TL;DR

MSWEP V3 delivers a global hourly precipitation dataset at $0.1^\circ$ resolution from $1979$ to present with about $2$-hour latency, achieved by stacking multiple ML models to generate baseline $P$ fields from diverse satellite and (re)analysis inputs and static predictors, followed by gauge-based corrections using an optimal interpolation scheme that accounts for gauge proximity, reporting times, and spatial correlation lengths. The approach includes a near-real-time MSWEP-NRT extension and a GPCC-based monthly correction with long-term bias adjustment, yielding a median daily KGE of $0.69$ that outperforms ERA5, IMERG-L, GSMaP, CHIRP, and prior MSWEP versions, with gauge corrections offering additional, regionally-variable improvements. The work demonstrates the value of multi-source ML baselines coupled with statistically grounded gauge corrections for stable, high-resolution precipitation estimation, enabling improved hydrological monitoring, forecasting, and risk management under a changing climate, and provides open access at www.gloh2o.org/mswep.

Abstract

We introduce Version 3 (V3) of the gridded near real-time Multi-Source Weighted-Ensemble Precipitation (MSWEP) product -- the first fully global, historical machine learning powered precipitation (P) dataset, developed to meet the growing demand for timely and accurate P estimates amid escalating climate challenges. MSWEP V3 provides hourly data at 0.1$^\circ$ resolution from 1979 to the present, continuously updated with a latency of approximately two hours. Development follows a two-stage process. First, baseline P fields are generated using machine learning model stacks that integrate satellite- and (re)analysis-based P and air-temperature products, along with static variables. The models are trained using hourly and daily observations from 15,959 P gauges worldwide. Second, these baseline P fields are corrected using daily and monthly gauge observations from 57,666 and 86,000 stations globally. To assess MSWEP V3's baseline performance, we evaluated 19 (quasi-) global gridded P products -- including both uncorrected and gauge-based products -- using observations from an independent set of 15,958 gauges excluded from the first training stage. The MSWEP V3 baseline achieved a median daily Kling-Gupta Efficiency (KGE) of 0.69, outperforming all evaluated products. Other uncorrected products achieved median daily KGE values of 0.61 (ERA5), 0.46 (IMERG-L V7), 0.38 (GSMaP V8), and 0.31 (CHIRP). Using leave-one-out cross-validation, the daily gauge correction was found to improve the median daily correlation by 0.09, constrained by the already strong baseline performance. We anticipate that MSWEP V3 -- accessible at www.gloh2o.org/mswep -- will enable more reliable monitoring, forecasting, and management of water-related risks in a variable and changing climate.

Figures (6)

• Figure 1: Flowchart of the ML model stack used to produce the hourly baseline MSWEP V3 $P$ fields.
• Figure 2: (a) Different ML model stacks were used for different periods and locations to account for spatiotemporal differences in the availability of dynamic predictors. The primary ML model stacks, associated dynamic predictors, and mean daily independent validation Kling-Gupta Efficiency (KGE; from Table\ref{['summary_performance_table_Global']}) are also shown. Note that alternative model stacks may be applied for a given period if any dynamic predictor is unavailable. (b) Conceptual illustration of how $P$ estimates from different ML model stacks are combined and harmonized with the reference (model_01) in the historical production pipeline.
• Figure 3: (a) Distance to the closest daily gauge that passed quality control and was used in the daily gauge-correction procedure ($n=57,666$; Appendix A "\ref{['DailyGaugeDataQC']}"). (b) $P$ correlation lengths based on the baseline MSWEP V3 (Appendix A, "\ref{['MethodsCorrelationLengths']}").
• Figure 4: Inferred reporting-time offsets (hours relative to midnight UTC) for daily gauges in (a) GHCN-D and (b) GSOD. For example, an offset of $-6$ h indicates that the daily total spans 18:00 UTC on the previous day to 18:00 UTC on the current day. Reporting times were inferred by maximizing agreement with the hourly MSWEP V3 baseline (Appendix A "\ref{['DailyGaugeDataQC']}").
• Figure 5: (a) Daily temporal correlation ($r_\textrm{dly}$) after gauge correction of the baseline versus distance to the nearest gauge. Each gray point ($n=57,666$) denotes a gauge for which post-correction $r_\textrm{dly}$ was evaluated using surrounding gauges (Appendix A "\ref{['MethodsDailyGaugeCorrection']}"). Solid lines show segmented regression fits to the post-correction $r_\textrm{dly}$ for each major Köppen-Geiger class beck_high-resolution_2023; dashed lines show fits to the baseline (uncorrected) $r_\textrm{dly}$. The $x$-axis tick labels indicate the breakpoints used in the segmented regressions. (b) Change in $r_\textrm{dly}$ after gauge correction ($\Delta r_\textrm{dly}$) versus distance to the nearest gauge. Each gray point ($n=57,666$) is the difference between post-correction and baseline $r_\textrm{dly}$; solid lines show segmented regression fits for each Köppen-Geiger class. (c) Global distribution of post-correction daily temporal correlation ($r_\textrm{dly}$) evaluated at gauge locations. (d) Global distribution of the change in daily temporal correlation caused by gauge correction ($\Delta r_\textrm{dly}$).