Evaluating the uncertainty in gridded rainfall datasets over Eastern Africa for assessing the model performance and understanding climate change In Support of: Planning for Resilience in East Africa through Policy, Adaptation, Research, and Economic Development (PREPARED) Project February 2016 1
1. Introduction Reliable observed rainfall data is crucial for evaluating the model performance and assessing and understanding climate change. In many parts of the world, however, observed rainfall records are often scarce, discontinuous and contain discrepancies. This is particularly true for many regions in Eastern Africa (Koutsouris et al. 2015). Although rain gauge datasets are available over the region, these products are not sufficient to capture rainfall at temporal and spatial scales. The lack of spatially and temporally continuous rainfall data over the region makes difficult in evaluating RCM simulations and understanding climate change. In the case of limited or no available station data, global satellite-gauge gridded datasets provide a viable alternative. Various satellite-gauge merged gridded datasets are available at different spatial and temporal resolutions. However, due to the quality of available station data and interpolation technique and blending methods used to combine satellite and gauage based products, the number of available observed datasets have a very wide spread at a particular space-time coordinate. As a result, the selection of reference data is becoming a challenge in model evaluation. Assessing climate change projections requires a careful evaluation of models ability to simulate the spatial and temporal scale of climate variables. In the past, different studdies used different observed rainfall datasets to evaluate/validate the regional model output and often proceed to assess projected changes in rainfal in the context of these observations. For example, Endris et al. (2013) used GPCC as a reference data for model evlaution over Eastern Africa. Kim et al. (2013) evaluated simulated rainfall data from a regional climate model against two different reference datasets (CRU & satellite data) and found that model performance substantially vary according to the reference data. Sylla et al. (2012) presented an intercomparison of different observed daily precipitation datasets and a validation exercise of RegCM3 model simulation, and found substantial discrepancies among the different observational datasets, and this makes it difficult to assess the model performance. In this report, the uncertainty between various observational datasets has been assessed over the Eastern Africa. This kind of analysis is very essential to highlight the differences in spatial and temporal rainfall estimates and also provide guidance to the choice of gridded rainfall data for assessing the model performance and understanding climate change over the region. 2
2. Data and methods 2.1 Data Six gridded rainfall datasets are analyzed and intercompared in this analysis. The CHIRPS- Blended data provided by ICPAC available at 5km spatial and monthly temporal resolution is used. CHIRPS-Blended (CHIRPS hereafter) incorporates satellite imagery with in-situ station data to create gridded rainfall time series and avilable starting 1981 to present for the whole GHA region. This data was first in binary (.bill) format, but now the data is available in netcdf format. Four gauge-based datasets are also used that are available at 0.5 o spatial and monthly temporal resolution: the Climatic Research Unit at the University of East Anglia (CRU TS3.10, 1901 2012), the Global Precipitation Climatology Centre (GPCC version 7, 1901 2009), the University of Delaware (UDEL version v3.01, 1901 2012) and Precipitation Reconstruction over Land (PRECL V3; 1948-2012). Moreover, the Global Precipitation Climatology Project (GPCP) which combines guage and staellite precipitation data available at 2.5 degree resolution is used eventhough the resolution is coarse. This data is avialabe 1979 to present at monthly temporal reolution. The table beow shows the six gridded observational datasets and their spatail and temporal resolutions. Table 1. List of gridded observational datasets used in this analysis. Product Spatial resolution Temporal resolution Type References CHIRPS- Blended 5kmx5km (0.05 o x0.05 o ) Monthly Satelite-gauge interpolated ICPAC CRU 0.5 o x0.5 o Monthly Gauge interpolated GPCC 0.5 o x0.5 o Monthly Gauge interpolated GPCP 2.5 x2.5 Monthly Satelite-gauge interpolated UDEL 0.5 o x0.5 o Monthly Gauge interpolated PRECL 0.5 o x0.5 o Monthly Gauge interpolated Harris et al. 2014 Schneider et al. 2015 Adler et al 2003 Legates and Willmott 1990 Chen et al. 2002 CHIRPS data is chosen as a reference field to evaluate the uncertainty in gridded observational datasets over the region. The choice of CHIRPS data is based on its high 3
resolution and also it incorporates many surface rainfall gauge data. 2.2 Methods Three statistical methods are used to quantify the uncertainty in gridded observational datasets. The three statistical measures considered are bias, correlation coefficient (R) and root mean square error (RMSE). Bias is a measure a systematic overestimation or underestimation of a data with respect to the reference data. Correlation measures similarity in temporal or spatial patterns between two datasets. Root Mean Square Error (RMSE) is a measure of the absolute mean difference between two datasets. The three statistics were calculated according to the following: R = 1 N N ( f n f )(r n r ) n=1 σ f σ r σ f 2 = 1 N N n=1 ( f n f ) 2 N σ 2 r = 1 (r n r ) 2 N n=1 n f n n=1 Bias = n n=1 n r n n=1 r n RMSE = n n=1 (r n f n ) 2 n where r n denotes estimates based on the reference data (i.e. CHIRPS), f n denotes the other gridded datasets being analyzed, n is the total amount of data pairs that were used in the analysis. In addition to these direct comparison statistics, mean precipitation amounts and standard deviations were also calculated as descriptive measures of each dataset. 4
3. Results and discussion 3.1 Seasonal mean The seasonl mean rainfall for the period of 1981-2010 is analysed for the six gridded datasets. Figure 2 represents the seasonal mean rainfall over GHA during JJAS (top panel), MAM (middle panel) and OND (bottom panel) as represented by different gridded observational datasets. All observation generally produce similar spatial pattern of rainfall with some discrepancies. In order to quantify the spatial agreement in seasonal mean rainfall, spatial correlation between the CHIRPS and the other gridded datasets was calculated. The spatial correlations are indicted in each plot in the bracket. The results indicate that there is a better agreement between CHIRPS and other observed datasets during JJAS and OND season compared to MAM season. Over all, GPCC and CRU have better agreement with CHIRPS than the other gridded observed datasets. PRECL show a week correlation with CHIRPS in all the three seasons. 3.2 Bias Figure 3 shows biases of rainfall from different observational datasets in reference to CHIRPS over the entire domain. GPCP and PRECL overestimated the rainfall over most of part of the region. In contrast, GPCC, CRU and UDEL underestimate the rainfall in most part of the areas. Although all the gridded datasets showed wet and dry biases in reference to CHIRPS, the two observation (CRU and GPCC) show smaller biases compared to the other datasets. 3.4 Annual cycle Figure 4 illustrates the annual rainfall cycle for the three homogeneous rainfall sub-regions. It is evident that all gridded observational datasets capture the main rainfall pattern in the three sub-regions clearly showing both the wet and the dry periods. The transition between the dry period and the wet period is also well captured in all observations. There is strong agreement among observations over NEA and SEA sub-regions. Over EEA (bimodal rainfall pattern), a discrepancy between observations in reproducing the rainfall pick and the timing of the transition from the wet to the dry period is observed. In reference to CHIRPS, all observations underestimate the OND rainfall pick. 5
3.5 Interannual variability The interannual variablity of seasnal mean rainfall over the three homogeneous rainfall subregions is shown in Figure 5. The temporal correlation between each observation and CHIRPS is computed and shown in Table 2. During JJAS over NEA, CRU and PRECL poorly represnted the interannual variability in comaprsion to CHIRPS with a correlation coefficent of 0.67 and 0.65, respectivelly. GPCC has a better agreemnt with CHIRPS in producing the interannual variablity of JJAS rainfall with correlation coefficent of 0.85. Over EEA during OND, CRU poorly reproduce the interannual variabilty. GPCC, GPCP and UDEL have better agreemnt with CHIRPS. In SEA, all observations except PRECL well repoduce the interannual variability of rainfall with a correlation coefficeint of > 0.95. Table 2. Correlation between seasonal mean rainfall between CHIRPS and the other gridded rainfall datasets during JJAS in NEA and OND in subregions EEA and SEA. Observation JJAS (NEA) OND (EEA) OND (SEA) CRU 0.67 0.71 0.95 GPCC 0.85 0.90 0.98 GPCP 0.79 0.93 0.97 UDEL 0.81 0.90 0.95 PRECL 0.65 0.88 0.89 3.6 Taylor diagram The performnace of the gridded observations on correlation (R), SD and RMSE can be visualized in Taylor diagram (Figure 6), where proximity to CHIRPS indicates good performance. Over NEA during JJAS, GPCC and UDEL had the best performance both with regard the correlation and standrd deviation. CRU shows high standard deviation compared to CHIRPS. Over EEA, GPCC and UDEL agrees best with CHIRPS with a standard deviation close to CHIRPS. GPCP overeastmate the variation, while CRU and UDEL underestimate it. Over SEA, all observations, except PRECL show a variation close to the reference data (CHIRPS). GPCC has relatively high correlation and low root mean square error (RMSE). This indicates that GPCC well captured the rainfall pattern in space and time. 6
4. Conclusions In this analysis, the performance of five gridded observational datasets (CRU, GPCC, GPCP, PRECL and UDEL) evaluated against the CHIRPS (5km resolution) over GHA. A number of performance measures were applied (Spatial correlation, temporal correlation, bias, root mean square error and standard deviation). The analysis indicated that discrepancies exit among the different observational datasets. It is also found that the performance largely depend on the season, sub-region and metrics used. GPCC and CRU have better spatial agreement with CHIRPS than the other gridded observed datasets. GPCC and GPCP well capture the interannual variability, while CRU and PRECL poorly reproduce the interannual variability on most of the sub-regions and season. Over all, GPCC had the best performance with regard to space and time analysis, and therefore can be used for the assessment of model performance over the region. 7
References Adler, R.F., G.J. Huffman, A. Chang, R. Ferraro, P. Xie, J. Janowiak, B. Rudolf, U. Schneider, S. Curtis, D. Bolvin, A. Gruber, J. Susskind, and P. Arkin, 2003: The Version 2 Global Precipitation Climatology Project (GPCP) Monthly Precipitation Analysis (1979- Present). J. Hydrometeor., 4,1147-1167. Chen, M., P. Xie, J. E. Janowiak, and P. A. Arkin, 2002: Global Land Precipitation: A 50-yr Monthly Analysis Based on Gauge Observations, J. of Hydrometeorology, 3, 249-266 Endris, H. S., Lennard, C., Hewitson, B., Dosio, A., Nikulin, G., & Panitz, H. J. (2015). Teleconnection responses in multi-gcm driven CORDEX RCMs over Eastern Africa. Climate Dynamics, 1-26. Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S.,... & Michaelsen, J. (2015). The climate hazards infrared precipitation with stations a new environmental record for monitoring extremes. Scientific data, 2. Harris, I. P. D. J., Jones, P. D., Osborn, T. J., & Lister, D. H. (2014). Updated high resolution grids of monthly climatic observations the CRU TS3. 10 Dataset. International Journal of Climatology, 34(3), 623-642. Kim, J., Waliser, D. E., Mattmann, C. A., Goodale, C. E., Hart, A. F., Zimdars, P. A.,... & Jack, C. (2014). Evaluation of the CORDEX-Africa multi-rcm hindcast: systematic model errors. Climate dynamics, 42(5-6), 1189-1202. Koutsouris, A. J., Chen, D., & Lyon, S. W. (2015). Comparing global precipitation data sets in eastern Africa: a case study of Kilombero Valley, Tanzania. International Journal of Climatology. Legates, D. R., & Willmott, C. J. (1990). Mean seasonal and spatial variability in gauge corrected, global precipitation. International Journal of Climatology,10(2), 111-127. Nikulin, G., Jones, C., Giorgi, F., Asrar, G., Büchner, M., Cerezo-Mota, R.,... & Sushama, L. (2012). Precipitation climatology in an ensemble of CORDEX-Africa regional climate simulations. Journal of Climate, 25(18), 6057-6078. Schneider, U., Becker, A., Finger, P., Meyer-Christoffer, A., Rudolf, B., & Ziese, M. (2011). GPCC full data reanalysis version 6.0 at 0.5: monthly land-surface precipitation from raingauges built on GTS-based and historic data. doi: 10.5676/DWD_GPCC. FD_M_V6_050. 8
Figures Figure 1: Map of the study area (GHA), with three subregions represented by boxes northern part of Eastern Africa (NEA), Equtorial part of Eastern Africa (EEA) and Southern part of Eastern Africa (SEA) that are utilized for the analysis. 9
Figure 2: Climatology of rainfall over eastern Africa during JJAS (top panel), MAM (middle panel) and OND (bottom panel) as represented by different gridded observational datasets. The spatial correlation between CHIRPS and other datasets is indicted in the bracket. 10
Figure 3: Biases of rainfall from different gridded observational dataset in reference to CHIRPS during JJAS (top panel), MAM (bottom panel) and OND (bottom panel). 11
Figure 4. Mean annual cycle of rainfall over NEA (top), EEA (middle), and SEA (bottom) from six different gridded observational datasets. 12
Figure 5. Interannual variability of rainfall during JJAS over NEA(top), during OND over EEA (middle), and during OND over SEA (bottom) from six different gridded observational datasets. 13
Figure 6: Taylor diagram displaying normalized statistical comparison of seasonal mean rainfall of various observed datasets in reference to CHIRPS over NEA during JJAS (top left), EEA during OND (top right), and SEA during OND (bottom left). 14