An optimized representation of precipitation in Jordan: Merging gridded precipitation products and ground-based measurements using machine learning and geostatistical approaches

Abstract

The accurate estimation of precipitation (P) at a high spatio-temporal resolution is vital in various applications such as climatic modelling, water resources management, drought and flood assessment, and climate change adaptation, among others. However, an adequate representation of P products in space and time remains challenging, particularly over regions with sparse or non-existent gauge-reference observations. Jordan, ranked among the top four driest countries in the world, urgently requires reliable P in good spatio-temporal resolutions to enable decision-makers and researchers to manage water resources effectively. In this study, seven state-of-the-art P products (MSWEPv2.8, ERA5, CHIRPSv2, CMORPHv1, PERSIANN-CDR, IMERG-FR, and ERA5 LAND) were evaluated against 124 gauge stations over the region using point-to-pixel evaluation at daily, monthly, annual, and seasonal temporal scales. Kling-Gupta efficiency as a continuous index with its three components (temporal dynamic r, bias ratio β, and variability ratio) was used to identify the systematic errors and uncertainties of the P products. Additionally, four categorical indices (probability of detection (POD), frequency bias (fbias), false alarm ratio (FAR) and the Critical success index (CSI) were used to assess the ability of the P products to capture different P intensities. The best performing daily scale P products were then resampled to a finer resolution of 0.05° (5 km) and merged with the gauge station observations to improve the representation of P over the region using two distinct approaches: i) machine learning approach, the Random Forest based MErging Procedure (RF-MEP), and ii) geostatistical approach, Kriging with External Drift (KED). We applied RF-MEP and KED over Jordan for the period 2001 - 2017 with a focus on its arid and climatic conditions; thus, we also applied the models to each climatic zone using daily observations of 80% of the gauge stations as a training dataset, and 20% were used for the verification of the merged P products. The results revealed that MSWEPv2.8 emerged as the top-performing P product. For this reason, and already being a merged dataset, MSWEPv2.8 was used as a benchmark in evaluating the merged products. For RF-MEP, The remaining datasets, excluding ERA5-LAMD and IMERGE-FR due to their poor performance, were merged with gauge observations, while KED was merged with the second-top performance product, ERA5. Both merged products demonstrated significant improvements in P patterns, linear correlation, bias, and variability at different temporal scales and in capturing different precipitation intensities. RF-MEP showed superior performance across Jordan compared to KED. However, KED outperformed RF-MEP in elevated terrains. Subsequently, a practical application of the newly merged P products was tested through simple drought assessment, using the Standardized Precipitation Index (SPI) specifically, SPI-12. The outcomes demonstrated that RF-MEP showed promising results in the detection of extreme long-term dry spells, highlighting its ability for practical application in drought assessment. Keywords: Precipitation; Gridded Precipitation products; Point to pixel evaluation; KGE; Categorical indices; Merging; RF-MEP; KED; SPI; Jordan