Muhammad Khalifa 1,2*, Wolfgang Korres 2, Samah Mahmood Saif 1,3, Nadir Ahmed Elagib 1, Oscar M. Baez-Villanueva 1,4, Mohammed Basheer 5, Saher Ayyad 1, Lars Ribbe 1, Karl Schneider 21 Institute for Technology and Resources Management in the Tropics and Subtropics (ITT), Cologne University of Applied Sciences, Cologne 50679, Germany2 Institute of Geography, University of Cologne, Albertus-Magnus-Platz, D-50923 Cologne, Germany3 Department of Sanitary and Environmental Engineering, University of Kassel, 34109 Kassel, Germany4 Faculty of Spatial Planning, Technical University of Dortmund University, Dortmund, Germany5 Department of Mechanical, Aerospace and Civil Engineering, The University of Manchester, Manchester, United Kingdom* Corresponding author: Muhammad Khalifa (msakhalifa@hotmail.com)Key Points:For efficiently managing water resources, precipitation needs to be monitored spatially and temporally.Public-domain Precipitation Products (PPs) are important sources of data, especially for data-scarce regions such as the Blue Nile Basin.We provided insights on the variations between 17 PPs using traditional evaluation methods and data mining techniques.AbstractThe efficient use of water requires understanding its spatial and temporal availability and pattern of use. However, in-situ measurements of the components of the hydrological cycle are often unavailable. This is particularly the case for precipitation. In this respect, Public-domain Precipitation Products (PPs) represent an alternative source of information. Nonetheless, precipitation estimates by PPs show discrepancies in spatial and temporal domains; thus, in-depth analyses of similarities and differences of these products is imperative to provide accurate precipitation estimations for water applications. We introduce and test a novel approach for evaluating the performance of PPs. This approach couples traditional evaluation methods (pixel-to-point and pixel-to-pixel) with data mining techniques (Hierarchical Clustering and Principal Component Analyses). It was used to assess the performance of 17 PPs over the Blue Nile Basin (BNB) for the period 2001-2005 on monthly and annual scales. A sensitivity analysis was carried out to test the affinity of the studied PPs. The analysis results were used to guide assimilating several PPs to create Merged Precipitation Products (MPPs). Results exhibit considerable differences between the studied PPs. Noticeable spatial and temporal discrepancies were found between the 17 PPs on the one hand and between PPs and rain gauge data on the other hand. Data mining techniques proved to be useful in detecting similar and dissimilar PPs. Given their advantages over traditional methods, these techniques should be used routinely in PPs assessment. The findings of the current research provide helpful insights to advance the use of PPs in water resources applications.Keywords: Precipitation, Blue Nile Basin, Public-domain, Hierarchical Clustering, Principal Component Analyses, Remote sensing.1 IntroductionWater monitoring is crucial for hydrological, ecological, and development purposes. Due to population growth and climate change, water has become increasingly scarce in many parts of the world (Kummu et al., 2016; Liu et al., 2017). Therefore, decision-makers are required to adopt immediate, efficient and sustainable management practices to meet current and future human development and environmental water demands. However, the lack of ground-based data is one of the main challenges that hinder good practices of water management (McDonnel, 2008). Effective management of water resources requires continuous monitoring and an accurate estimation of the spatio-temporal patterns of different components of the hydrological cycle such as precipitation, evapotranspiration, runoff, and water storage changes (Ayyad et al., 2019; Cosgrove & Loucks, 2015; Fernández-Prieto et al., 2012; Su et al., 2010; Sun et al., 2018). In most regions of the world, water availability is directly linked to precipitation amount and seasonality (Dinku et al., 2007; Ligaray et al., 2015; Noy-Meir, 1973). Variations in the spatio-temporal patterns of precipitation can cause environmental hazards such as floods and droughts, which have direct socio-economic impacts (Brown & Lall, 2006), and often result in loss of lives and infrastructure. For example, Masih et al., (2014) reported that during 1965-2012, drought events affected ~67 million people over Ethiopia, bringing an estimated economic loss of above 92 million US$ and a death toll of more than 400,000. These numbers emphasize the need for accurate precipitation data to support decision-making, especially in areas vulnerable to high climate variability such as the Nile Basin (Bastiaanssen et al., 2014; Beyene et al., 2010; Cao et al., 2018b).Traditionally, precipitation has been measured using in-situ rain gauges (Gabriele et al., 2017; Kidd, 2001). However, the accuracy in the characterization of precipitation, when only ground-based measurements are used, depends largely on the density and distribution of the rain gauge network (Shaghaghian & Abedini, 2013). While radar data can provide a spatially distributed estimation of precipitation (Yoon et al., 2012), rain gauges are considered the most reliable source of precipitation measurements at the point scale (Villarini et al., 2008) and they still required for calibration and validation purposes. Despite that, the rain gauges are sensitive to environmental conditions (Michelson, 2004), and the accuracy of their records has to be controlled (Levy et al., 2017). However, in many regions (especially in developing countries), rain gauges are sparsely distributed (Kaba et al., 2014), and their number is decreasing (Sun et al., 2018). Rain gauges are sensitive to environmental conditions (Michelson, 2004), and the accuracy of their records needs to be checked (Levy et al., 2017). A dense network is expensive and hard to maintain, hindering an accurate spatial representation of the precipitation patterns, especially in high altitude areas. Systematic under-catch of gauge measurements (Beck, Wood, McVicar, et al., 2019), unsystematic errors such as gaps in time series (Woldesenbet et al., 2017), latency in data availability, in addition to inaccessibility of data are additional challenges that limit the use of rain gauge data in many regions of the world (Montesarchio et al., 2015; Thiemig et al., 2012). In Africa, the implementation of an adequate rain gauge network is challenging because of driver factors such as the desired accuracy and the cost of implementation, maintenance and data collection (Pardo-Igúzquiza, 1998).The recent technological development in sensors, algorithms and new satellite missions designed to measure environmental processes, have enabled the opportunity to derive gridded precipitation estimates. This has enabled the opportunity to account for the spatial distribution of precipitation (Kidd, 2001; Zambrano-Bigiarini et al., 2017), thus providing data which is otherwise often not feasible to obtain. The public-domain policy of these products has encouraged the use of their datasets for different applications such as drought assessment (Agutu et al., 2017; Gao et al., 2018; Sahoo et al., 2015; Zambrano-Bigiarini et al., 2019), flood forecasting (Tekeli et al., 2017), hydrological modeling (Kite & Pietroniro, 1996; Siddique-E-Akbor et al., 2014), water balance studies (Bastiaanssen et al., 2014; Karimi, et al., 2013), among others.A wide range of sensors, input data, and estimation algorithms are used to produce these Public-domain Precipitation Products (PPs) (Sun et al., 2018). Accuracy of the PPs estimation can be affected by climatological of catchment-specific factors such as elevation (Ayehu et al., 2018; Dinku et al., 2018; Habib et al., 2012; Hirpa et al., 2010). Hence, the accuracy of the PPs in representing the spatio-temporal precipitation patterns varies greatly depending on the region. Although some studies have reported an overall accuracy as high as 95% (Karimi & Bastiaanssen, 2015), the accuracy of these PPs might vary at different temporal scales and geographic settings (Baez-Villanueva et al., 2018).Traditionally, the PPs are evaluated by comparing their estimates with in-situ measurements, for example, (i) using a pixel-to-point analysis, where the rain gauge data are compared to the estimates of the respective grid-cells of the PPs (Bai & Liu, 2018; Burton et al., 2018; Cao et al., 2018; Gebrechorkos et al., 2018; Thiemig et al., 2012); (ii) using a pixel-to-pixel approach, which compares a gridded version of the rain gauge data with the corresponding grid-cell of the PPs product (Amitai et al., 2009; Bajracharya et al., 2015; Chen et al., 2014; Saber et al., 2016). Additionally, the evaluation of PPs could be carried out indirectly by using the PPs to force a hydrologic model and evaluate the simulated discharge with streamflow observations (Beck et al., 2017; Casse et al., 2015; Chintalapudi et al., 2014; Tramblay et al., 2016; Voisin et al., 2007). In some cases, there are no enough ground-based data to evaluate these products; and therefore, information on the performance of the different PPs can be assessed through a cross-correlation analysis (Salih et al., 2018). Such inter-comparison would provide the relative differences in precipitation between the different PPs, and might shed some light on their similarities and differences. Given the large number of data that needs to be handled in such evaluation approaches using grid-cells values, data mining techniques, such as Hierarchical Clustering Analysis (HCA) and Principal Components Analysis (PCA) can be effective in reducing effort and time needed to assess these big data (Lever et al., 2017; Zhang et al., 2017).Since all PPs have advantages, limitations, and uncertainties, merging different PPs may provide a better estimation for precipitation (Baez-Villanueva et al., 2020; Bastiaanssen et al., 2014; Peña-Arancibia et al., 2013; Xie et al., 2003). However, most of the available ensemble algorithms are complex to implement, and their performance normally improves when increasing the number of rain gauges (Baez-Villanueva et al., 2020). Therefore, merging products over extremely data-scarce regions such as the Nile Basin remains a challenge. To this end, simple merging methods, like the one followed in the current research, would benefit massively from the comprehensive assessment that couples traditional evaluation methods with data mining techniques suggested herein.Many previous studies have been conducted to evaluate the performance of PPs over the Blue Nile Basin (BNB) (e.g. Abera et al., 2016; Mekonnen and Disse, 2018; Romilly and Gebremichael, 2011). For a summarized review of some of these studies, the reader is referred to the supplementary material (Table S1). These studies have focused only on the upstream part (UBNB) of the basin, with few exceptions that targeted the lower BNB (LBNB) (e.g. Basheer et al., 2018), and only a limited number of PPs were evaluated. These studies have evaluated the PPs products performance through a direct comparison with rain gauge data to assess their ability to represent the precipitation patterns (Thiemig et al., 2012). This approach is limited over data-scarce regions because there is no enough data to implement an informative evaluation (Bastiaanssen et al., 2014). It is worth to mention here that the number of rain gauges in operation over the BNB is decreasing (see the supplementary material: Text S1; Figures 1 and 2).Therefore, the objectives of the current research are: (i) to detect the similarities and differences between 17 PPs over the BNB at monthly and annual temporal scales using mean annual precipitation values through a pixel-to-pixel inter-comparison and data mining techniques, and to cluster them into groups based on their similarities; (ii) to evaluate the performance of these PPs over the BNB using rain gauge data; and (iii) to evaluate the applicability of this integrated analysis of PPs in guiding simple merging procedures of PPs. The merging exercise aims at creating Merged Precipitation Products (MPPs) to improve the precipitation estimation, as a potential solution to improve precipitation estimation of PPs over data-scarce regions. The present analysis aims to advance the current understanding of the performance of PPs over the BNB, as an example of data-scare regions. To the best of our knowledge, such a comprehensive investigation at the given scale integrating traditional evaluation approaches with data mining techniques has not been conducted so far, neither for the BNB nor any other river basins.2 Data and Materials2.1 Study areaThe BNB is a transboundary river basin shared by Ethiopia and Sudan (Fig. 1a). The basin has an area of about 307,177 km2, of which around two-thirds are located in Ethiopia (Upper BNB, UBNB) and the rest is in Sudan (Lower BNB, LBNB). Whereas the UBNB is characterized by complex topography, the LBNB is relatively flat. The BNB contributes to nearly 62% of the total streamflow of the Nile River (Amdihun et al., 2014), and is crucial for food and hydropower production (Allam & Eltahir, 2019; Elagib et al., 2019; Wheeler et al., 2016). The rainfed and irrigated agricultural schemes in the basin produce a large fraction of the annual domestic food production of Ethiopia and Sudan (Awulachew et al., 2012; Elagib et al., 2019). The BNB is the main source of water for the Gezira irrigation scheme - one of the largest irrigated schemes in the world with an area of around 0.88 million hectares (World Bank, 1990). Precipitation in the BNB is difficult to predict (Cheung et al., 2008; Meze-Hausken, 2004) and highly variable in time and space (Beyene et al., 2010; Conway, 2000). The intra-annual and inter-annual variability of precipitation has a direct impact on rainfed agriculture, and also on irrigated agriculture as a result of reduced river flows under drought conditions (Kim et al., 2008; Siam & Eltahir, 2017). The rainy season in the basin is relatively short and lasts for only five months (from June to October). The mean annual precipitation in the basin varies from ~120 mm at the outlet of the basin in Khartoum (in the LBNB) to more than 2000 mm in some parts of the UBNB (Roth et al., 2018).
