How Skilful Are Global Precipitation Datasets Over Saudi Arabia Region for Spring Rainy Season?

ABSTRACT For Saudi Arabia, it is crucial to decide relatively more useful precipitation products for Climate research, developing applications, climate model verifications and more recently for developing AI/ML models. We used freely available 12 datasets: CRU, GPCC, CPC, CHIRPS, ERA5, IMERG, GSMaP, PERSIANN, CMORPH, MSWEP, GPCP, CMAP, to examine skill of these monthly datasets over the Kingdom of Saudi Arabia, covering a major portion of the Arabian Peninsula for the Spring Season (MAM). Spatial analysis and analysis with respect to station data for five climatic zones were carried out. Different skill scores like mean, bias, variance, standard deviation, correlations, POD, FAR and CSI were examined. It is observed that different datasets have different categories of spatial biases and variability in mean rain across the Arabian Peninsula. There is uncertainty in the capability among datasets in capturing the inter‐annual variability for the AP region. This makes these datasets potentially unreliable for long‐term climate trend analysis in this region. Every single dataset failed to accurately capture precipitation in the orographic highlands zone, underestimating the values, indicating a fundamental limitation of current global algorithms in handling complex terrains in arid regions. GSMaP and CMORPH: both multi‐satellite products were noted for overall poorer skill and an inability to represent the region's precipitation accurately. IMERG and GSMaP, both in spite of being from the same GPM mission, performed drastically differently in skill scores, with IMERG emerging as superior across most Saudi Arabian climatic zones. For five climatic zones, the names of datasets to be avoided for applications are summarised in a tabular form. The ensemble means of two types of data capture well the mean, bias, CC and SD compared to all other datasets. It is recommended to use a selective ensemble mean for research and applications by removing datasets with lower skills.