Evaluation of CHIRPS Satellite Gridded Dataset as an Alternative Rainfall Estimate for Localized Modelling over Uganda ()
1. Introduction
Precipitation is one of Uganda’s most critical and valuable resources. Its availability and variability have direct implications on the country’s overall development given the predominance of rain-fed agriculture [1]. Indeed, rain-fed agriculture and agriculture-related production are vital to the Ugandan economy; they averagely employ 80% of the labour force, account for 42% of the GDP, and bring in 90% of export earnings [2] [3]. However, Uganda is in the deficit of a dense rain gauge network and long-term in situ precipitation measurements are required for detailed assessment of her rainfall [4] [5]. The rain fall records are typical of non-continuity with only a few years of quality data, sparse rain gauge network, incomplete, inconsistent and non-reliable datasets. Notably, understanding the dynamics of climate variations and extremes that are consistent with the region requires long-term accurate precipitation data representations and a dense network of in situ observation stations [6] [7]. The World Meteorological Organization (WMO) set the classical period of assessing climate variability to 30 years.
Research [8] puts the optimum rain gauge network density between 14 km2/gauge and 38 km2/gauge, with an average of 25 km2/gauge. With Uganda having a total area of 241,550 km2, this puts the total gauge requirement per 25 km2 close to about 9500 weather stations. However, as of 2013, Uganda had 21 weather stations [9]. With the proliferation of automated weather stations, as of 2021, the total number of weather stations had been improved to 37. Incidentally, of these, 25 are working, 7 were vandalized and 5 had power problems. This puts the total landmass covered to 625 km2 out of the total 241,550 km2, accounting for only 0.0026% gauge coverage. This has characterized the country with a critically sparse and unreliable rain gauge network. Traditionally, it should be noted that set rain gauges are affected by animal and human interference, uncalibrated measuring mechanisms, level placement, wind, poor monitoring and evaporation [10] [11] [12]. This further worsens the dependency on completeness and consistency of in situ measurements from rain gauges. Also, relative to satellites that can provide precise measurements at every location, for developing countries, the sparsely distributed weather stations are characterized by poor coverage for rural areas and require interpolation between stations [13] to relay general coverage of a region. This makes the reliance on satellite data sets inevitable to achieve alternative estimates of rainfall with adequate sampling density, accuracy, completeness and reliability. However, satellite-based precipitation estimates are often not assessed for integration into operational and decision-making applications due to the lack of empirical research on their uncertainty and reliability [14]. For Uganda, no explicit countrywide empirical research has been done to assess the suitability of satellite gridded precipitation datasets as an alternative rainfall estimate at a local scale. It should be noted that the region is challenged with scarce and inconsistent precipitation data. The objective of this research was, therefore, to evaluate satellite gridded datasets as an alternative rainfall estimate for modelling rainfall at a local scale. The research will enable satellite-based precipitation estimates to be integrated into the operational analysis of rainfall dynamics as a proxy for in situ rainfall data in Uganda.
2. Satellite Gridded Datasets as Alternative Rainfall Estimates
Numerous satellite gridded precipitation data sets exist at a global and regional scale [15]. Though satellites have trouble measuring some ground phenomena such as precipitation as compared to weather stations, they provide a more complete spatial coverage of various parameters over a landscape [13]. This has prompted research to understand the contrast between the two sets of data. Throughout the literature, several studies have compared satellite remote sensing products to in situ data [16] [17] [18] [19] [20]. The satellite-based datasets have generally presented a better performance at annual and seasonal scales. Some scholars [21], explored the performance of stochastic models for spatial rainfall downscaling to reproduce statistics for precipitation observable at a local scale. The analysis, however, would not allow preference among the studied models. Other scholars [22] have explored algorithms of blending the in situ measurements with satellite observations to derive more accurate products for local rainfall analytics. However, none of the research has been conclusive leaving a gap in the analysis of local precipitation using satellite gridded rainfall estimation datasets.
3. Materials and Methods
3.1. Study Area
This research was carried out within the spatial domain of the Republic of Uganda. Uganda is a landlocked country in the eastern part of Africa and lies within the northern and southern hemispheres, being crossed by the Equator. Neighboring countries are Kenya to the East, South Sudan to the North, the United Republic of Tanzania and Rwanda to the South and the Democratic Republic of Congo to the West. According to [23], the country has a total surface area of 241,550 km2 of which 41,743 km2 (17.2%) is occupied by open water and swamps, and 199,807 km2 is open land. Uganda is a plateau surrounded by four main mountain ranges; namely: Rwenzori in the west (with a peak of 5110 m), Elgon in the East, Mufumbira in the Southwest, and Moroto in the Northeast. The lowest part is within the Albert Nile at 620 m above sea level. This topographic complexity is associated with a high density of micro-environments where climatic variables become difficult to study [24]. In essence, this makes the area suitable for this study relative to other gauge complexities. The map in Figure 1 shows the location of Uganda in Africa, its administrative districts and large water bodies.
3.2. Rainfall Data Sets
3.2.1. Satellite Gridded Data Sets
The Climate Hazards Group InfraRed Precipitation with Stations (CHIRPS) data set was used as the proxy for satellite gridded datasets. The datasets blend station precipitation data with infrared Cold Cloud Duration (CCD) at 0.05˚ * 0.05˚ (latitude, longitude) spatial resolution [25]. It has been adopted for this research mainly for three reasons: 1) It has high sub-regional suitability in comparison to other precipitation satellite products [26] [27] [28]; 2) It has substantial improvement over non resampled satellite products due to gauge based bias correction [29]; 3) Other alternative satellite products offer a coarse spatial resolution; Global Precipitation Climatology Project (GPCP) and the Climate Precipitation Centre (CPC) Merged Analysis of Precipitation (CMAP) have a 2.5˚ * 2.5˚
spatial resolution while the African Rainfall Climatology version 2 (ARC2) and Tropical Rainfall Measuring Mission (TRMM) have a spatial resolution of 0.1˚ * 0.1˚ and 0.25˚ * 0.25˚ respectively.
3.2.2. In Situ Measurements
In Situ precipitation data was gotten from Uganda National Meteorological Authority (UNMA). The data was for multiple weather stations. However, the datasets were assessed for completeness, continuity, consistency, and reliability. On this basis, only 12 of the 25 data stations were utilized. Their distribution is illustrated in Figure 2. Though long time series are often required, the datasets used in this research were for the temporal period of 2012 to 2020. Beyond this period there were no sufficient datasets that fulfilled the criteria suitable for effective comparison with the gridded datasets.
4. Statistical Metrics
Spatial Inference research on precipitation studies has utilized a vast continuum of statistical criteria to assess agreement between models [24] [27] [28]. This research adopted the statistical metrics shown in Table 1.
Figure 2. Location map of the in situ precipitation stations.
Table 1. Adopted statistical metrics.
5. Methodology (Figure 3)
Datasets collected from the weather stations were assessed for consistency, completeness and continuity for the study period between 2012 and 2020. The data were interpolated between stations across the country. The validity of the resulting interpolation was assessed through independent weather stations (15% of the original weather stations) repetition were used to validate the Interpolation. The geostatistical ordinary Kriging was used as previous studies showed that it achieved the best results in precipitation interpolation [24] [33]. A fishnet was generated over Uganda with a 5 km point grid. Buffers of 5 km to 20 km from the in situ measurement stations were generated to clip the points of the fishnet. The clipped points were used for the extraction of values from the interpolated surface and satellite gridded dataset (CHIRPS). A bivariate linear regression was run between the extracted values of the interpolation surface and the CHIRPS surface in iteration to acquire the best sample points extraction range for comparing the two surfaces at an R2 greater than 0.75. This was in consideration of [8] analysis on optimal distance for in situ station coverage. Buffers of 5 km and 20 km were chosen as the furthest and nearest ranges for further analysis of the comparison between the CHIRPS satellite gridded dataset and in situ station measurement rainfall estimates (the minimum 5 km range was chosen because the CHIRPS has a 0.05˚ * 0.05˚ (latitude, longitude) spatial resolution). These were subjected to further spatial statistical metrics shown in Table 1 to derive a detailed evaluation of the relation between the two surfaces.
6. Results and Discussion
Precipitation data from the in situ weather stations was mapped by Kriging
interpolation for each year as shown in Figure 4. The results showed an average representation of most precipitation being received along the northwest-southeast belt and the Lake Victoria basin. From the results, it can also be deduced that the southwest-northeast belt received less rainfall. This is consistent with observations of the cattle corridor that show a pronounced increase in dry spells along the southwest-northeast corridor [34]. Related research [35] also observed the highest rainfall in the north-west and central east of Uganda and Lake Victoria vicinity
Satellite gridded datasets were collected for the same epoch (2012-2020) as the in situ data sets and mapped as shown in Figure 5. Similar to observations by [35] the lowest precipitation totals are in the northeast (Karamoja region) and the south-west while the highest levels of precipitation are observed in the Lake Victoria vicinity, central-east and north-west.
Research [8] approximates optimal rain gauge coverage to about 25 km2 beyond which the quality of representation drops. From the in situ measurement positions in Figure 2, buffers points at 5 km intervals were generated as shown in Figure 6 for extraction of values from each surface at minimum 5 km and maximum 20 km ranges.
Between the in situ and satellite observations at 5 km, the coefficient of determination (R2) showed very strong collinearity of 0.91 as shown in Figure 7. The Nash-Sutcliffe Efficiency (NSE) coefficient = 0.88, Percentage of Bias (PBias) = −0.24, and the RMSE-observations Standard deviation Ratio (RSR) = 0.35.
Figure 4. Interpolated average annual precipitation.
Figure 5. Average annual precipitation from satellite gridded data sets—CHIRPS.
Figure 7. Collinearity between in situ and satellite precipitation observations within a 5 km radius of the weather station.
Between the in situ and satellite observations at 20 km, the coefficient of determination (R2) showed strong collinearity of 0.75 as shown in Figure 8. The Nash-Sutcliffe Efficiency (NSE) coefficient = 0.461, Bias Percentage (PBias) = 0.84, and the RMSE-observations Standard deviation Ratio (RSR) = 0.51.
Several studies [36] [37] have adopted the performance rating of the statistical metrics R2 [30], and RSR, NSE, PBias [31] shown in Table 2. According to this rating, at 5 km there is very good collinearity between the in situ data and the CHIRPS satellite model with an R2 of 0.91. This implies that a high proportion of the variance in the in situ observations is accounted for in the CHIRPS gridded satellite observations. The RSR of 0.35 low value at 5 km indicates a very good approximation by the CHIRPS satellite model. The 0.88 NSE also implies a very good simulation of in situ data by the CHIRPS satellite data. The PBias of −0.24 indicates minimal deviation of the data being evaluated which also accounts for the very good simulation by the CHIRPS model. At 20 km an R2 of 0.75 indicated a good collinearity between the in situ and satellite observation. The RSR of 0.51 also indicated a good approximation of the in situ data by the CHIRPS satellite model. The PBias of 0.84 also indicated an acceptable deviation between the in situ data and the gridded satellite data. However, the NSE of 0.461 indicated an unsatisfactory simulation of the in situ data by the CHIRPS satellite model.
Similar statistical metrics were run for each of the in situ stations and its 5 km and 20 km radius as shown in Table 3. Within the 5 km radius, 75% of the stations presented with an R2 above 0.86 indicating very good collinearity between the in situ observations and CHIRPS satellite data model. However, the Kabale station presented an unsatisfactory R2 of 0.44. Also, 60% of the stations indicated an above satisfactory RSR statistic indicating a good approximation of the in situ data by the CHIRPS satellite data model. Correspondingly, 65% of the stations presented an NSE statistic above satisfactory indicating an acceptable simulation of in situ data by CHIRPS satellite gridded data. All the stations had above satisfactory PBias indicating acceptable deviation between the two datasets. At 20 km, only 25% of the records presented R2 and RSR above satisfactory levels indicating low levels of collinearity and approximation between the models, 40%
Figure 8. Collinearity between insitu and satellite precipitation observations within a 20 km radius of the weather station.
Table 2. Performance rating of the statistical metrics.
Table 3. Statistical metrics at 5 km and 20 km sample range of in situ weather stations.
of the stations presented an NSE statistic above satisfactory level implying that more than 60% of the in situ observation couldn’t be interpreted at this range. However, the PBIAS indicated acceptable deviations between the two models.
7. Conclusion
The assessments and monitoring of precipitation variation, changes and extremes are critical to an agriculturally based economy. However, for Uganda, the availability of consistent, complete, and reliable precipitation data required for this monitoring and assessment has remained a challenge. This research sets out to validate the use of CHIRPS satellite gridded data sets as an alternative source of precipitation data. Based on the [8] review of best rainfall in situ observations being in the optimal range of 25 km2, this research forms an evaluation basis to validate the use of CHIRPS satellite gridded dataset as an alternative precipitation estimation dataset by comparing the 5 km range and radius sample points between surfaces of interpolated in situ surfaces and CHIRPS satellite dataset. It further shows that widening the validation area beyond the optimal coverage of the rain gauge generates dwindling effects on the validation between the in situ and CHIRPS satellite gridded dataset surfaces. Spatially based statistical metrics that include the coefficient of determination (R2), the Nash-Sutcliffe Efficiency coefficient (NSE), Bias Percentage (PBias), the RMSE-observations and the Standard deviation Ratio (RSR) were used to check collinearity, simulation, deviation and approximation respectively between the two rain estimation models. General results indicate that the R2 = 0.91, NSE = 0.88, PBias = −0.24 and RSR = 0.35. This attests that the CHIRPS Satellite gridded datasets provide a good approximation and simulation of in situ station data with high collinearity and minimum deviation. This can be attributed to the spatial resolution and temporal resolution of the CHIRPS dataset as well as its gauge bias correction. The results tally with related studies [27] in other regions that have found CHIRPS datasets superior to interpolation surfaces and sparse rain gauge data in the comprehensive estimation of rainfall. With a 0.05˚ * 0.05˚ (latitude, longitude) spatial resolution, CHIRPS satellite gridded rainfall estimates are able to provide rainfall representation at a local scale. Essentially these results reward research science in regions like Uganda that have sparse rain gauges networks characterized by incomplete, inconsistent and unreliable data with an empirically researched alternative source of rainfall data. It further provides a platform to interrogate the rainfall dynamics at a local scale in order to infuse local policy with evidence-based formulation and application.