Key points are not available for this paper at this time.
Accurate precipitation (P) estimates are crucial for a wide range of applications, including water resource management, disaster risk reduction, agricultural planning, and infrastructure development. Over the past few decades, numerous gridded P products have been developed, with varying temporal and spatial resolutions, derived from diverse data sources, and employing different methodologies and algorithms. However, these products frequently exhibit significant uncertainties, errors, and biases, underscoring the importance of selecting the most suitable product for each application. In this study, we conducted a comprehensive evaluation of the strengths and weaknesses of over 20 freely available global gridded P products. We used the European RADar CLIMatology (EURADCLIM) gauge-radar dataset, the US Stage-IV gauge-radar product, and observations from approximately 20,000 global stations as ground truth. Our assessment included several new products, such as PDIR-Now and GPM+SM2RAIN, as well as an experimental Random Forest (RF) model, a potential new version of the Multi-Source Weighted-Ensemble Precipitation (MSWEP) product. For the assessment, we employed a broad range of performance metrics sensitive to various aspects of P time series, including the versatile Kling-Gupta Efficiency (KGE) and its components (correlation, bias, and variability), as well as the Critical Success Index (CSI), wet day bias, peak bias, and trend error. Additionally, we assessed the relative performance in different physiographic regions, seasons, and P regimes, and among various product types (satellite, (re)analysis, gauge, and combinations thereof). The RF model showed the best overall performance, achieving a mean CSI of 0.42. In comparison, the current MSWEP version, CHIRP, ERA5, GSMaP and IMERG achieved mean CSI values of 0.40, 0.21, 0.36, 0.32, and 0.32, respectively. Our study highlights the stark differences in performance among various state-of-the-art P products and provides a baseline for the development of new machine learning-based P products.
Wang et al. (Mon,) studied this question.