Seasonal Prediction of Summer Precipitation over East Africa Using NUIST-CFS1.0

Temesgen Gebremariam ASFAW and Jing-Jia LUO

1Institute for Climate and Application Research (ICAR)/CICFEM/KLME/ILCEC, Nanjing University of Information Science and Technology, Nanjing 210044, China

2Institute of Geophysics Space Science and Astronomy, Addis Ababa University, Addis Ababa 1176, Ethiopia

ABSTRACT East Africa is particularly vulnerable to precipitation variability, as the livelihood of much of the population depends on rainfed agriculture. Seasonal forecasts of the precipitation anomalies, when skillful, can therefore improve implementation of coping mechanisms with respect to food security and water management. This study assesses the performance of Nanjing University of Information Science and Technology Climate Forecast System version 1.0 (NUISTCFS1.0) on forecasting June-September (JJAS) seasonal precipitation anomalies over East Africa. The skill in predicting the JJAS mean precipitation initiated from 1 May for the period of 1982-2019 is evaluated using both deterministic and probabilistic verification metrics on grid cell and over six distinct clusters. The results show that NUIST-CFS1.0 captures the spatial pattern of observed seasonal precipitation climatology, albeit with dry and wet biases in a few parts of the region. The model has positive skill across a majority of Ethiopia, Kenya, Uganda, and Tanzania, whereas it doesn’t exceed the skill of climatological forecasts in parts of Sudan and southeastern Ethiopia. Positive forecast skill is found over regions where the model shows better performance in reproducing teleconnections related to oceanic SST. The prediction performance of NUIST-CFS1.0 is found to be on a level that is potentially useful over a majority of East Africa.

Key words: East Africa, seasonal precipitation forecasts, probabilistic verification, ensemble prediction

1. Introduction

Precipitation variability has important socioeconomic impacts on East Africa. Precipitation extremes (floods and droughts) severely affect the economies of East Africa,primarily through their impacts on agriculture production and pastoralism. The recent droughts in 2011, 2014, and 2015 in Ethiopia and the drought in 2017 in Somalia,Kenya, and parts of Ethiopia and surrounding countries had devastating consequences (Viste et al., 2013; Philip et al.,2018; Kew et al., 2021). Climate conditions in the first half of 2011 are claimed to have been the driest in the past 60 years for some regions of Somalia, northern Kenya, and southern Ethiopia (FEWS NET, 2011), causing severe food shortages for about 11 million people.

The seasonal cycle of rainfall over East Africa is influenced by the latitudinal migration of the Inter-Tropical Convergence Zone (ITCZ), which follows the seasonal variations of solar insolation and is strongly modulated by the region’s diverse topography. In addition, various influences by large-scale teleconnections lead to complex spatiotemporal patterns of rainfall (Nicholson, 2017). The migration stages of the ITCZ can be separated into four phases (Seregina et al., 2019, 2021): two positions during boreal summer (June-September, JJAS) and winter (December-March,DJFM), and two periods in boreal spring (March-May,MAM) and autumn (October to December, OND). The northern and northwestern parts of eastern Africa (i.e., central and northern Ethiopia, Eritrea, Djibouti, northern Uganda,and South Sudan) receive the bulk of their precipitation during JJAS boreal summer season (Williams et al., 2012; Nicholson, 2017). For instance, this season accounts for nearly 50% to 80% of the annual rainfall over Ethiopia’s agricultural regions (Korecha and Barnston, 2007).

Previous studies have identified the north-south displacement of upper-level jet features, zone of maximum convection, and southern boundary of the thermal low over northern Africa as the primary drivers of the interannual and longer time-scale variability of East African boreal summer precipitation (Nicholson, 2017); all of which appear to be influenced by tropical sea surface temperature (SST) anomalies (Camberlin, 1995, 1997; Riddle and Cook, 2008;Kucharski et al., 2009; Segele et al., 2009a, b; Diro et al.,2011a). Other studies have found very strong correlations between East African summer rainfall and the strength of the Indian monsoon, independent of a mutual dependence on El Ni?o-Southern Oscillation (ENSO; Camberlin, 1997;Vizy and Cook, 2003). The main moisture source for JJAS seasonal precipitation in East Africa originates from the tropical Atlantic Ocean (specifically the Gulf of Guinea), the rainforest region of the Congo Basin, and the Indian Ocean (Camberlin, 1997; Riddle and Cook, 2008; Williams et al., 2012;Viste and Sorteberg, 2013a, b; Viste et al., 2013). Moisture from the Gulf of Guinea appears to be the least important but the most variable from year to year (Nicholson, 2017).

Tremendous effort has been devoted to establishing predictive relationships between East African summer seasonal rainfall and remote indicators. A majority of these studies examined the predictability of JJAS rainfall with SSTs in the tropical Pacific and Indian Oceans as predictors (Camberlin, 1995; Gissila et al., 2004; Segele and Lamb, 2005;Korecha and Barnston, 2007; Segele et al., 2009a; Diro et al., 2011a, b). They found that the interannual variability of East African rainfall during this season is primarily governed by the remote influence of ENSO but that local factors near Africa and in the Atlantic and Indian Oceans play a role as well. These statistical linkages between East African rainfall variability and tropical SSTs have been used by the Greater Horn of Africa Climate Outlook Forum(GHACOF) and national weather services to issue operational seasonal forecasts (Mason and Chidzambwa, 2008;Viste et al., 2013; Mason et al., 2019; Walker et al., 2019)over several decades. These GHACOF consensus forecasts were found to have positive skill, but forecast probabilities for the near-average category were found to be systematically too high, indicating a tendency to “hedge” to average conditions (Mason and Chidzambwa, 2008; Walker et al.,2019).

Recently, attention has been paid to the use of dynamical models in operational seasonal forecasting in this region (Graham et al., 2012; Mwangi et al., 2014; Walker et al., 2019). Several studies have been carried out to evaluate the skill of different products in seasonal precipitation predictions (e.g., Ogutu et al., 2017; MacLeod, 2018; Shukla et al., 2019; Walker et al., 2019; Young and Klingaman,2020). However, the target areas in most of the previous studies were limited to Equatorial East Africa (Kenya, Uganda,and Tanzania) and were limited to analyzing precipitation forecast skill for MAM and OND seasons. In these two seasons, predictions of statistical models are generally more skillful than those of dynamical models. However, dynamical models do outperform statistical models in some cases(Young and Klingaman, 2020). For instance, Walker et al.(2019) found a dynamical model to have higher skill for predicting East African rainfall than statistical models.

This paper aims to examine the skill of Nanjing University of Information Science and Technology Climate Forecast System version 1.0, referred to as NUIST-CFS1.0 hereafter, in predicting JJAS seasonal precipitation over East Africa. While the GHACOF consensus forecasts are issued for MAM, JJAS, and OND seasons based on a mix of dynamical and statistical forecast systems, our verification focuses on JJAS because it is the primary rainy season for the northern two-thirds of the study area (central and northern Ethiopia, Eritrea, Djibouti, northern Somalia, South Sudan,and Sudan; Fig. S1 in the electronic supplementary materials, ESM), which includes regions with large populations that are growing rapidly (Pricope et al., 2013). The monsoon over this region exhibits high interannual variability(e.g., Beltrando and Camberlin, 1993; Camberlin, 1997;Segele et al., 2009a), and failure of the JJAS precipitation has caused severe droughts in the region (e.g., Korecha and Barnston, 2007). NUIST-CFS1.0 is a fully coupled atmosphere-ocean global climate model (Luo et al. 2003,2005a). It has good performance in simulating and predicting both ENSO and Indian Ocean Dipole (IOD), including their magnitudes, periods, and the spatial distribution of the SST anomalies in the Indo-Pacific region (Luo et al.,2005b, 2007, 2008a, b). The skill evaluation is performed using both deterministic and probabilistic scores. The representation of ENSO and IOD and their teleconnections in NUIST-CFS1.0 are also examined. Section 2 provides the data and methods, and section 3 presents the results. The summary and discussion are given in section 4.

2. Data and methodology

2.1. Verification data

The global land-only Climate Hazards Group InfraRed Precipitation with Station (CHIRPSv2) rainfall dataset(Funk et al., 2015a) was used as observational data in this study. It incorporates numerous National Meteorological Service gauge records and uses multiple satellite observations and a sophisticated model to interpolate rainfall to a 0.05° resolution over the period from 1981 to near-present. CHIRPS is built based on a global 0.05° monthly precipitation climatology (CHPclim). The CHPclim bias over different parts of the world, including Ethiopia and the Sahel, is low (～3% or less). CHPclim also appears to perform well in data-sparse regions with complex terrain (Funk et al., 2015a, b). Compared to similar satellite rainfall products, the quality of CHIRPS over East Africa is significantly better. Specifically, CHIRPS has higher skill and lower bias than the African Rainfall Climatology version 2 (ARC2), and it is also slightly better than the Tropical Applications of Meteorology using Satellite data version 3 (TAMSAT3) product at decadal and monthly time scales (Dinku et al., 2018).

CHIRPS data is available at 0.05° and 0.25° horizontal resolutions and can be downloaded from https://data.chc.ucsb.edu/products/CHIRPS-2.0/. In this study, bilinear interpolation is used to match the observation with the NUISTCFS1.0 grid.

To investigate influences of the tropical oceans, SST data from the Extended Reconstructed Sea Surface Temperature, Version 5 (ERSSTv5; Huang et al., 2017) is used.Wind and geopotential data are obtained from ERA5 (Hersbach et al., 2020).

2.2. Seasonal forecast system

NUIST-CFS1.0 is built based on a fully coupled atmosphere-ocean global climate model. The atmospheric component (ECHAM4.6; Roeckner et al., 2003) has T106 resolution (about 1.1° × 1.1° latitude-longitude grid mesh) with 19σ-pressure vertical levels. The oceanic component (OPA8.2;Madec et al., 1998) has the resolution of a 2° Mercator mesh(increased to 0.5° in the latitudinal direction near the equator) with 31 vertical levels. Both model coupling physics and initial conditions are perturbed separately in three different ways to constitute an ensemble of nine members. To generate realistic and atmosphere-ocean well-balanced initial conditions required for the hindcasts, observed weekly NOAA OISST values are assimilated into the coupled model with strong restoring strengths. Details are given in Luo et al. (2005a, b, 2008a) and He et al. (2020).

We verify the NUIST-CFS1.0 seasonal forecasts for the period of 1982-2019. The forecasts are initialized on the first day of every month in each year and run out to 24 months. This study focuses on forecasts with a 1-month lead time, referring to a forecast for June-September that is initialized on 1 May.

2.3. Clustering

As the precipitation over East Africa exhibits high spatial variability (as discussed in section 1), it is useful to divide the region into clusters of homogeneous precipitation climate. Clusters are identified by applying the k-means clustering algorithm (e.g., Straus et al., 2007; Dawson et al.,2012; Dawson and Palmer, 2015; Yang et al., 2020) to observed monthly precipitation after interpolating to the NUIST-CFS1.0 grid (i.e., 1.1° × 1.1°). Similar data points are partitioned into corresponding k clusters through repeated iterative operations. This partition is constructed so that the ratio of variance between cluster centroids to the average intra-cluster variance is maximized. This condition corresponds to desiring cluster centroids to be far apart and for the points within each cluster to be close together. The silhouette coefficient (a measure of cluster cohesion and separation) of differing cluster sizes (2-20) and prior knowledge of the East Africa rainfall distribution is used to decide the number of clusters.

As a compromise between representing all climate zones and avoiding having a cluster number that is too large for analysis, a total of six clusters was selected (Fig. 1). All time-series analyses in this study are performed based on the six clusters, which are briefly described here in terms of their climatology.

Two distinct types of rainfall classes are found in East Africa (Fig. 2). Many regions show a unimodal rainfall distribution with a maximum in JJAS being observed over Sudan, South Sudan, and central and northern Ethiopia (i.e.,Cluster 1, Cluster 3, and Cluster 5). A distinct bimodal rainfall pattern with two maxima occurring in spring and late autumn is observed in equatorial East Africa, including Kenya, Tanzania, Rwanda, Burundi, the southern part of Somalia, and south-eastern Ethiopia (i.e., Cluster 2, Cluster 4, and Cluster 6). The highest rainfall amounts occur over Cluster 5, while the lowest amounts occur over Cluster 1.

Fig. 1. (a) Topographic elevation map (m) of East Africa and (b) clusters of homogeneous rainfall zone, indicated by distinct colors. The cluster analysis is carried out with the k-means clustering algorithm based on the observed(CHIRPS) monthly precipitation.

Fig. 2. Observed (CHIRPS) monthly climatology of precipitation (mm month-1) for the different clusters (see Fig. 1),with the bars in the center representing the mean and the shaded areas showing the distribution.

2.4. Verification method

The skill of NUIST-CFS1.0 in forecasting the JJAS seasonal precipitation over East Africa is evaluated using several deterministic and probabilistic evaluation metrics. The verification was carried out for 1-month lead prediction of JJAS seasonal precipitation (i.e., forecasts initiated from 1 May for the target season of JJAS).

For the deterministic forecasts based on the nine-member ensemble mean, anomaly correlation coefficient (ACC)and root-mean-square error (RMSE) are used to assess the prediction skill. ACC and RMSE are two of the most widely used skill metrics for assessing seasonal climate forecast performance (e.g., Luo et al., 2008; Doblas-Reyes et al., 2013;Mishra et al., 2019; Kim et al., 2021). ACC assesses the degree of linear correspondence between the forecast anomalies and the anomalies of the observed climate variable.And RMSE of the ensemble mean prediction measures the distance between the forecast magnitude and the observed magnitude. The Relative Operating Curve Skill Score (ROCSS)and the Ranked Probability Skill Score (RPSS) are also widely used metrics for evaluating the skill of probabilistic forecasts (Goddard et al., 2003; Palmer et al., 2004; Ogutu et al., 2017; Walker et al., 2019). The ROCSS and the RPSS for upper, middle, and lower tercile forecasts are determined after converting the forecasts into forecast probabilities, assuming that each ensemble member prediction is an equally probable forecast. At each grid point, three categories are defined by sorting and partitioning the CHIRPS observations and NUIST-CFS individual member forecasts, with each category having equal occurrence frequency in the assessment period. The baseline probability for any category is therefore 33.3%, with each category expected to occur, on average, once in three years. The lowest third of the data values are defined as below normal, the middle third of the values are near normal, and the upper third of the values are above normal.

The ROC score measures the ability of the forecast to discriminate between events and non-events (Mason, 1982). A ROC score for each tercile category (above, near, and below normal) can be obtained from the area under the ROC curve. To construct the ROC curve, thresholds of predicted probabilities for each tercile category (above, near, and below normal) are first determined. The ROC curve is then generated as a plot of the hit rate against the false alarm rate for these thresholds. A ROC score may be transformed into a ROCSS:

where A is the area under the ROC curve.

The RPSS measures the improvement of multi-category forecasts relative to a reference forecast (the sample climatology) and considers the probability for each category.The ranked probability score (RPS) represents the sum of the squares of the difference between the categorical cumulative forecast probabilities and the corresponding observed categorical cumulative probability. For a given probabilistic forecast and observation pair, RPS is given by the following equation (Weigel et al., 2007; Wilks, 2011):

where J is the number of categories (here it is 3), Yfcrefers to the relative occurrence frequency of ensemble members in the corresponding category, and Oobsrepresents the observation probability in the category (Oobsis equal to one if the observed data falls into the jth category, and it becomes zero otherwise). For a perfect forecast, the value of the RPS is zero.

The RPSS is defined in terms of RPS as follows:

The ROCSS and RPSS compare the skill of a forecast to that of a standard reference (i.e., the climatological forecast and observed climatology, respectively), such that zero means the forecast is as good as the reference. Positive values imply an improvement, and negative values imply worse skill than the reference (Weigel, 2012).

The generalized discrimination score (GDS), which is a measure of how well the forecasts are able to discriminate among varying observations (Mason and Weigel, 2009; Weigel and Mason, 2011), is used to indicate whether the forecasts may be potentially useful. According to Ziervogel et al. (2005), forecasts need to be correct at least 60%-70% of the time in order to be of use for smallholder farmers, which corresponds to a GDS of 0.6-0.7. Hence, forecasts with a GDS below 0.6 are considered potentially useless, and forecasts with a GDS above 0.7 are useful, even for small-scale applications (e.g., Gubler et al., 2020). The GDS is computed for ensemble forecasts with continuous observations as follows (Weigel and Mason, 2011; Weigel, 2012):

Consider a set of n observations x1, ···, xnand corresponding ensemble forecasts, ··· ,. Let there be m ensemble members, and letbe the ith ensemble member of the tthforecast. The GDS is then given by:

The GDS of a forecast that does not contain any useful information is 0.5, which corresponds to random guessing.The more successfully the forecasts are able to discriminate the observations, the closer the score is to 1.

2.5. ENSO and IOD indices

The performance of NUIST-CFS1.0 in predicting the relationship of ENSO and IOD with precipitation in East Africa is assessed. The impact of ENSO (e.g., Trenberth, 1997) on precipitation in East Africa is analyzed using the SST anomalies in the Ni?o-3.4 region (5°N-5°S, 120°-170°W). The SST anomalies in the Ni?o-3.4 region exceeding ±0.5°C are used to classify El Ni?o and La Ni?a events. The IOD index is calculated as the difference between the Western Tropical Indian Ocean (WTIO; 10°N-10°S, 50°-70°E) and the Southeastern Tropical Indian Ocean (SETIO; 0°N-10°S,90°-110°E) SST anomaly, as was defined by Saji et al.(1999). The IOD index exceeding ±0.4°C is used to identify positive and negative IOD events.

Composite analysis techniques are performed to investigate potential influences of ENSO and IOD on the precipitation in East Africa. Moreover, a partial correlation technique (Yule, 1907; Timm and Carlson, 1976) is used to show an independent relationship between the two variables while excluding influences arising from the other variable. For example, the partial correlation of East African precipitation with the IOD, in which the influence of ENSO is removed (Yamagata et al., 2004; Behera et al., 2005;Bahaga et al., 2019; Walker et al., 2019), is defined as follows:

where r13is the correlation between IOD and precipitation,r12is the correlation between IOD and ENSO, and r23is the correlation between ENSO and the precipitation. Similarly,the partial correlation of East African rainfall with ENSO,independent of the IOD’s influence, is calculated.

3. Results

3.1. Deterministic skill

Figure 3 displays the climatology of JJAS seasonal precipitation based on the NUIST-CFS1.0 prediction and the observations as well as the biases of the model prediction.The mean precipitation of the CHIRPS observations shows rainfall maximums over the western and central highlands of Ethiopia, South Sudan, near the border with Central Africa, and Congo (Fig. 3b). In Sudan, the maximum precipitation, ranging from 400 to 900 mm, is located over its southern part, while the remainder of the country receives around 100 mm of rainfall during the JJAS season. The large-scale precipitation distribution across East Africa is captured well by NUIST-CFS1.0 (Fig. 3a), although the strong precipitation extends too far into Sudan and Eritrea. A climatological mean wet bias is exhibited over relatively large swaths of northern East Africa (Fig. 3c). The seasonal mean wet bias is particularly large in the normally dry regions of Sudan and northern and eastern Ethiopia, reaching up to 400 mm. In contrast, the NUIST-CFS1.0 prediction underestimates the seasonal rainfall by 300-400 mm with a large dry bias over the maximum precipitation regions of western Ethiopia and South Sudan. The dry bias also extends to the southwestern part of East Africa, including Uganda, the southern part of Kenya, and Tanzania.

The ensemble-mean skill of NUIST-CFS1.0 in predicting the JJAS mean precipitation anomalies during the period of 1982-2019 over East Africa is assessed based on the ACC and RMSE metrics (Fig. 4). A significant positive correlation between the forecasted and observed precipitation anomalies during the period of 1982-2019 is apparent over several parts of East Africa, including Ethiopia, South Sudan, Uganda, Kenya, and northern Sudan. The ACC skill is positive across most of East Africa, with particularly high values (over 0.6) in northeastern and southwestern Ethiopia,parts of Kenya, South Sudan, and Uganda. These high ACCs are statistically significant at the 95% confidence level. However, the AAC skill is negative over southeastern Ethiopia, southern Sudan, and northern Tanzania, which indicates a model deficiency in predicting interannual variability of precipitation over these areas. In terms of the RMSE,it is shown that NUIST-CFS1.0 performs well over equatorial East Africa. However, it exhibits large error over most portions of Sudan, Somalia, and northwestern and southeastern Ethiopia. Generally, areas of larger forecast errors correspond to areas of negative ACCs.

Fig. 3. East Africa JJAS seasonal precipitation climatology (mm season-1) during 1982-2019 based on (a) the nine-member ensemble mean of NUIST-CFS1.0 forecast initiated from 1 May, (b) the CHIRPS observations, and (c) the difference between the NUIST-CFS1.0 prediction and the observation (i.e., the model prediction bias).

Fig. 4. (a) Anomaly correlation coefficient (ACC) and (b) root-mean-square error (RMSE) based on the ensemble mean forecast of JJAS mean rainfall anomalies during 1982-2019 using NUIST-CFS1.0. Dots in (a) on each grid indicate significant positive correlation at the 5% significance level.

3.2. NUIST-CFS1.0 prediction over the six climate zones

Seasonal rainfall patterns over East Africa are very complex due to the existence of complex topography, including large inland water bodies, rift valleys, and snow-capped mountains (e.g., Indeje et al., 2000; Seregina et al., 2021).The interactions of large-scale atmospheric forcing with the regional heterogeneous topography give rise to dramatic variations in the spatial distributions of the JJAS climatological mean rainfall (Sun et al., 1999a, b). In addition, seasonal rainfall anomalies in the region tend to have a coherence that is confined across small subregions (e.g., Ogallo et al., 1988;Ogallo, 1989; Indeje et al., 2000; Mutai and Ward, 2000).This suggests a necessity to perform a verification over homogeneous rainfall regions (clusters) over East Africa. Historically, homogeneous rainfall regions were used in verification of the numerical climate model simulations over the region (e.g., Sun et al., 1999a, b; Diro et al., 2011b, c;Tsidu, 2012). In this study, cluster analysis is performed using k-means clustering, as described in section 2.3, to identify six homogeneous precipitation regions. The evaluation of NIUIST-CFS1.0 seasonal forecasts over these distinct clusters are then presented using scatter and anomaly correlation plots.

The performance of NIUIST-CFS1.0 in predicting the seasonal precipitation varies from cluster to cluster. Figure 5 shows scatterplots of the ensemble-mean forecasts versus the corresponding observations of the seasonal precipitation during 1982-2019 for below, near, and above normal tercile categories over each cluster (Fig. 1). For all clusters,except Cluster 1 and Cluster 3, most of the scatter points are around the diagonal line, indicating that the ensemble mean forecasts show a high skill in reproducing the observed seasonal precipitation. However, the ensemble-mean forecast overestimates the seasonal precipitation over Cluster 1(encompasses semiarid areas of Sudan, Eritrea, northeastern Ethiopia, Somaliland, and coast of Somalia) and Cluster 3 (consists of the central highlands of Ethiopia, the southern part of Sudan, and South Sudan). The ensemble-mean forecast slightly underestimates the precipitation over Cluster 6. Further comparison of the three categories reveals that these overestimations are relatively high during the below-normal precipitation years.

Figure 6 displays the skill in predicting the interannual variations of areal averaged seasonal precipitation anomalies over each cluster. The JJAS anomalies are computed for each ensemble member and the ensemble-mean by removing the seasonal climatology of the corresponding ensemble member and the ensemble-mean, respectively. Similarly,the observed anomalies are computed by removing the climatology of the seasonal mean precipitation. The violin plot in Fig. 6 reveals the ensemble member spread of the seasonal precipitation anomaly predictions. The NUIST-CFS1.0 ensemble spread generally encompasses the observed seasonal mean precipitation anomalies over the six clusters for most years during 1982-2019, except 1984, 1988, 1990,1996, 2008, 2009, and 2015 for Cluster 3 and 1988, 1996,1998, 2003, 2009, and 2015 for Cluster 5. Note that 2009 and 2015 are El Ni?o years and 1988 and 1998 are La Ni?a years. A temporal anomaly correlation coefficient skill(between the ensemble-mean predictions and the observations) of 0.37 is achieved over Cluster 2, followed by 0.34 and 0.31 over Cluster 3 and Cluster 6, respectively. The model is able to correctly predict the sign of the precipitation anomaly for many of the extreme years, especially over Cluster 2 and Cluster 6.

3.3. Probabilistic forecast skill

The overall performance of NUIST-CFS1.0 in delivering a probabilistic categorical forecast is assessed by ROCSS, RPS, and GDS. The ROCSS provides the skill of NUIST-CFS1.0 in forecasting precipitation for the three categories (below, near, and above normal). The RPSS and GDS summarize the model prediction scores over the three tercile categories.

Figures 7a and b show a spatial distribution of the ROCSS for the upper and lower tercile categories. The ROCSS ranges from 0 to 1; any value higher than 0 indicates a forecast is better than that from the climatology forecast. For both categories, there is a coherent region of positive skill over Kenya, Uganda, South Sudan, Somaliland, Djibouti, and most parts of Ethiopia. This suggests that the model’s ability to discriminate the categories is better than a random forecast over those areas. However, the model forecast shows lower skill over most parts of Sudan and southeastern Ethiopia. The model prediction skill is relatively better for the wet category than the dry category over northern Sudan and northeastern Ethiopia. However, in Eritrea, Somaliland, and the coast of Somalia, the model has better skill in predicting the dry category.

The RPSS and GDS in predicting JJAS precipitation are presented in Figs. 7c and d. The RPSS measures the ability of a forecast system to capture the proximity between the forecast and the observed probability; scores above zero indicate superior skill relative to the climatological probabilistic forecast. In general, the RPSS of NUIST CFS1.0 is positive over a majority of East Africa, particularly over the central highlands of Ethiopia, the southeastern portion of South Sudan, most parts of Eritrea, Somaliland, Uganda, and Kenya. In terms of the probabilistic skill measured by the RPSS, the forecasts are generally skillful over the regions where the ROCSS is higher (cf. Figs. 7a-c).

The usefulness of the NUIST-CFS1.0 forecast is further demonstrated with the spatial distribution of the GDS(Fig. 7d). As was mentioned before, GDS is a measure of how well the forecasts are able to discriminate between varying observations, and a score greater than 0.5 indicates that the forecast contains useful information. The result shows that the model forecast can convey useful information in most parts of East Africa. The highest score appears over Ethiopia, South Sudan, Kenya, and Uganda. The GDS over these areas is greater than 0.7, implying that, in 70% of the cases, the forecasts are able to correctly discriminate between continuous observations. Conversely, the forecast appears to have no use in the southern part of Sudan, northwestern and southeastern Ethiopia, and eastern Tanzania.

Fig. 5. Scatter plots of observed (CHIRPS) and the ensemble mean forecast of JJAS mean precipitation (mm season-1) using NUIST-CFS1.0 for near-normal (black dot), above-normal (green dot), and below-normal (red dot)years during 1982-2019 for each grid point in the clusters. Terciles are defined at each grid point.

For a more quantitative assessment, percentages of grid cells with positive skill scores are displayed in Fig. S2 in the ESM. For the ROCSS and RPSS, scores above zero that indicate enhanced skill relative to the climatological probabilistic forecast are taken to calculate the fraction of areas where the forecast is skillful. The ROCSS percentages for both the lower and upper tercile category are above 60%,with a slightly higher value for the latter. The RPSS is greater than zero on about 50% of all the grid cells. For the GDS, a score greater than 0.5 is considered in calculating the percentage of the grid cell. At least 65% of all grid cells have a GDS of above 0.5; this implies that useful information can be obtained from the forecast over most parts of East Africa for each cluster.

Fig. 6. Time series of JJAS precipitation anomaly (mm season-1) predicted from 1 May based on NUIST-CFS1.0 ensemble mean (green dots), ensemble members spread (violin plot), and the CHRIPS observations (red dots) for each cluster during 1982-2019. Anomaly correlation coefficients (ACC) between the ensemble mean forecast and the CHIRPS are shown at the top of each panel.

Fig. 7. (a and b) The Relative Operating Curve Skill Score (ROCSS) in predicting (a) upper tercile category and (b)lower tercile category of JJAS seasonal precipitation. (c) Ranked Probability Skill Score (RPSS) and (d) Generalized Discrimination Score (GDS) in predicting JJAS seasonal precipitation tercile categories. Only areas of positive skill(i.e., ROCSS ＞0, RPSS ＞0, and GDS＞0.5) are shown in colors, and areas of no skill are masked in gray.

Fig. 8. Correlation of JJAS seasonal precipitation over East Africa with Ni?o-3.4 SST index (upper panels) and the IOD index (bottom panels) based on (a, c) the NUIST-CFS1.0 forecasts and (b, d) the CHIRPS observations.

Fig. 9. Correlations of the observed SST (ERSST) anomalies with JJAS seasonal precipitation over each cluster based on the NUIST-CFS1.0 forecasts (left column) and the observations from CHIRPS (right column).

3.4. Prediction of remote impacts of ENSO and IOD

ENSO is one of the principal modes influencing climate at the seasonal scale in East Africa (e.g., Indeje et al.,2000; Camberlin and Philippon, 2001; Camberlin et al.,2001; Endris et al., 2019), and the predictability of ENSO and its teleconnections have been identified as the main source of predictability at the seasonal scale in East Africa(e.g., Segele et al., 2009; Diro et al., 2011a, b; Nicholson,2014, 2017). Previous studies have shown the relationship between JJAS seasonal precipitation in East Africa and SST anomalies in the equatorial Pacific, particularly highlighting the importance of ENSO. MacLachlan et al. (2015) and Arribas et al. (2011) further demonstrated that a good representation of ENSO and its teleconnections is required for skillful regional climate prediction on seasonal time scales. In contrast, the relationship of JJAS seasonal precipitation in East Africa with IOD has not yet been well documented.Recently, anomalously enhanced rainfall during a positive IOD event in the JJAS season was noted, especially in the Sahel and tropical Africa, by Preethi et al. (2015). They further suggested that a co-occurrence of a positive IOD event,in a linear sense, could reduce the impact from the tropical Pacific drivers.

The precipitation aggregated zonally from 20°E to 50°E during La Ni?a, El Ni?o, and neutral years clearly indicates the influence of ENSO on JJAS seasonal precipitation in East Africa (Figs. S3a and b in the ESM). Consistent dry and wet conditions are observed during El Ni?o and La Ni?a years, respectively. This indicates the existence of a relationship between East Africa JJAS seasonal precipitation and ENSO. NUIST-CFS1.0 is able to predict this relationship well in lower latitudes, albeit with a slight underestimation. However, the model fails to differentiate the observed variations between El Ni?o and La Ni?a years in higher latitudes north of about 12°N. A similar analysis is conducted to assess the skill in predicting the influence of IOD on JJAS seasonal precipitation over East Africa (Figs. S3c and d in the ESM). The observed precipitation differences among the positive, neutral, and negative IOD events are not as large as the ENSO counterparts. The model shows lower prediction skill in discriminating the precipitation conditions among the three phases of IOD.

Skill in predicting the impacts of ENSO and IOD on JJAS precipitation over East Africa is further investigated based on the forecasts initialized from 1 May by differentiating El Ni?o years from La Ni?a years (Figs. S4a and b in the ESM) and by differentiating negative IOD years from positive IOD years (Figs. S4c and d in the ESM), respectively.Figures S4a and b present the composite precipitation differences (El Ni?o-minus-La Ni?a) based on the model forecasts (left panels) and the observations (right panels). Overall, the model predicts well the difference between El Ni?o and La Ni?a years over most parts of East Africa. The impacts of ENSO, such as the dry conditions over Ethiopia,Uganda, and Kenya, are captured well. However, the model has deficiencies in predicting the observed dry conditions over Sudan and the wet conditions over eastern coastal regions. Comparison of composite precipitation differences based on the model forecasts and the observations for the IOD case (positive IOD-minus-negative IOD) indicates that the model predicts well the wet conditions over eastern parts of East Africa, including Somalia, Somaliland, and the eastern part of Ethiopia (Figs. S4c and d in the ESM).However, the model fails to predict the dry conditions over South Sudan, the southern part of Sudan, and northeastern Ethiopia.

To successfully predict the JJAS seasonal precipitation,a model must correctly predict both the evolution of ENSO or IOD and the atmospheric response to the predicted SST anomalies (as was discussed in section 2.5). Figure 8 illustrates the partial correlations of JJAS seasonal precipitation on each grid of East Africa with the Ni?o-3.4 SST index and the IOD index based on the NUIST-CFS1.0 forecasts and the CHIRPS observations. The spatial pattern of the partial correlation with the Ni?o-3.4 index is reproduced well by the NUIST-CFS1.0 forecasts over most parts of Ethiopia, Kenya, Uganda, Rwanda, and Burundi. However,the model fails to predict the observed correlations over Sudan, Somalia, and southeast Ethiopia. This may contribute to the low predictive skill over these areas. The region where the observed partial correlation is represented well by the model coincides with the region of the highest forecast skill (recall Figs. 4 and 7). In addition, the IOD teleconnections are also predicted well over most parts of East Africa,except parts of Somalia, South Sudan, and northeast Ethiopia.

Significant geographical variations have been observed in the influences of the large-scale teleconnections over East Africa (e.g., Ogallo et al., 1988; Segele et al., 2009a). This suggests the need to examine the performance of the model in predicting the relationship between the SST anomalies and the precipitation over the individual clusters. Figure 9 shows the correlations of the tropical SST anomalies at each grid with the predicted (left panels) and the observed (right panels) precipitation over each cluster. Both sets of panels display a clear large-scale pattern, with the strongest and most coherent correlations appearing in the Pacific, representing the copious influences of ENSO-related SST anomalies.The model predicts the negative correlations of precipitation with the tropical Pacific SST in all the clusters, except for Cluster 1 and Cluster 6. However, the correlations predicted by NUIS-CFS1.0 are of a smaller magnitude than the observations, suggesting that the remote impacts of the tropical SST variations may be too weak in the model predictions. The correlations of the tropical SST anomalies with the precipitation over Cluster 1 is not predicted well by the model. This is consistent with the result that the negative and positive seasonal precipitation anomalies during El Ni?o years (e.g., 1982 and 2015) and La Ni?a years (e.g.,1988 and 2010) are not predicted well (recall Fig. 6a). For Cluster 6, in contrast, the model predicted correlations are stronger than the observed, consistent with the result that the negative and positive precipitation anomalies over Cluster 6 during these El Ni?o and La Ni?a year are overestimated in the model forecasts (recall Fig. 6f).

Modelling experiments by Vizy and Cook (2001, 2003)and Diro et al. (2011a) emphasized the importance of air pressure and circulations over the Atlantic and southwestern Indian Oceans, including westerly low-level wind from the Gulf of Guinea and the East African Low Level Jet(EALLJ).

Composite anomalies for 850-hPa winds and geopotential height during anomalous wet and dry-precipitation years over Cluster 5 are shown in Fig. S5 in the ESM for both observation (ERA5) and 1-month lead NUIST-CFS1.0 prediction. A normalized JJAS precipitation anomaly spatially averaged over the cluster exceeding ±1 is used to define wet and dry years during the period of 1982-2019.

Prominent observed low-level wind features in wet/dry composites are the westerly/easterly winds from the Atlantic and the southwestern Indian Oceans, which are associated with the EALLJ. Wet/dry composites are also associated with observed negative/positive geopotential height anomalies over the Indian Ocean and strong/weak St.Helena (North Atlantic) and Azores (southeastern Atlantic)highs.

The composites from NUIST-CFS1.0 captured the observed low-level winds features. However, the wind anomalies are weaker in magnitude, which may affect the moisture transport from the Atlantic and southwestern Indian Oceans. There are also clear limitations in predicting the geopotential height anomalies, including the strong St. Helena and Azores highs during the wet years and the cyclone/anticyclone anomalies over the Indian Ocean during wet/dry years.

The forecasts of JJAS SST indices in the tropical Indian Ocean and Pacific Ocean initialized from 1 May are compared with the observations and persistence forecast in Fig. 10. The persistence forecast uses the average SST anomalies over the entire month of April. The model predicts well the variations of the observed Ni?o-3.4 SST index in most years during 1982-2019 with a high anomaly correlation coefficient of 0.85. A similar level of predictive skill is found in the WTIO (correlation skill of 0.79) and the SETIO (correlation skill of 0.75). However, the model prediction underestimates the magnitude of the SETIO SST anomalies. This may contribute to the relatively low correlation skill of 0.45 in predicting the IOD index, which is lower than the ACC of the persistence forecast (Fig. 10d). This result suggests that further efforts are required to improve IOD prediction to allow for better forecasts of the JJAS precipitation in East Africa.

Fig. 10. Time series of the NUIST-CFS1.0 ensemble mean forecasts and the observations (ERSST) of JJAS SST anomalies of (a) Ni?o-3.4, (b) Western Tropical Indian Ocean (WTIO), (c) Southeastern Tropical Indian Ocean (SETIO), and (d) Indian Ocean Dipole (IOD) index. ACC skills for NUIST-CFS1.0 and persistence forecasts are shown at the right bottom of each panel.

4. Summary and discussion

Seasonal climate prediction is of great importance in East Africa where climate variability affects socioeconomic activities such as agriculture. In this study, the evaluation of NUIST-CFS1.0 skill in predicting JJAS seasonal precipitation in East Africa is presented. The climatological mean precipitation of the NUIST-CFS1.0 forecast is compared with the observed climatology, and cluster analysis is used to partition the result into different homogenous precipitation regimes. NUIST-CFS1.0 forecast skill is assessed using both deterministic and probabilistic metrics, including ACC, RMSE, ROCSS, RPSS, and GDS. The skill of the model is assessed separately for each grid cell and temporally over six clusters based on the forecasts initiated from 1 May.

NUIST-CFS1.0 realistically predicts the spatial patterns of the observed seasonal climatology, albeit with a dry bias in the southern part of Sudan and South Sudan and a wet bias over the western tip of Ethiopia and Uganda.NUIST-CFS1.0 prediction skill varies from region to region. The ACC skill is positive across most of East Africa. Negative ACC skill over some parts of this study area, such as Eastern Sudan and the southern part of Tanzania, was also found in the European Centre for Medium-Range Weather Forecasts (ECMWF) precipitation forecasts for June-August evaluated by Ogutu et al. (2017), while the positive ACC skill over Kenya and southeastern and eastern Ethiopia is higher in NUIST-CFS1.0. Assessment based on the ROCSS and RPS reveals that categorical precipitation forecasts are skillful and potentially useful, except for in a few areas such as Sudan and northeastern and southeastern Ethiopia. For most parts of the study area, skill in terms of the RPSS is similar to that found by Shukla et al. (2019)for 1-month lead forecasts of July-September using the North American Multi-Model Ensemble (NMME). NUSTCFS1.0 skill is better for central Ethiopia and Somaliland,but the NMME is superior over northwest Ethiopia and South Sudan. The ROCSS for upper and lower terciles is also found to be more or less in agreement with the results of Ogutu et al. (2017) for bias corrected ECMWF forecasts.Ogutu et al. (2017) found higher skill over a large part of East Africa (particularly over central Ethiopia, Kenya, and Uganda) compared to that found by Arribas et al. (2011)and MacLachlan et al. (2015) for 1-month lead forecasts using the UK Met Office Global Seasonal forecast system(GloSea). Nevertheless, GloSea skill from https://www.metoffice.gov.uk/research/climate/seasonal-to-decadal/gpc-outlooks/glob-seas-prob-skill shows better skill over northeast Ethiopia, South Sudan, and the southwestern part of Sudan.High precipitation prediction skill is found over central Ethiopia, Uganda, and Kenya. However, low precipitation prediction performance is found over Sudan and southeastern Ethiopia. It is worth noting that positive probabilistic skill is found over most grid points of East Africa; this suggests that the model has positive skill in predicting large-scale precipitation features in East Africa. Few other studies have analyzed seasonal precipitation forecast skill over this particular region and forecast period (JJAS). The observation data used by previous studies also vary. However, NUIST-CFS has been shown to have comparable skill with forecast systems which are usually evaluated and used to some extent for this area.

Correlations of the tropical SST anomalies with precipitation over the six clusters in East Africa based on the NUIST-CFS1.0 forecasts are compared to the observations.The result shows that the model replicates the major SST patterns linked with the precipitation, except for Cluster 1 and Cluster 6. In general, skill is better for clusters where the link of the precipitation with the oceanic surface conditions is well captured by the model.

The results indicate that the JJAS seasonal precipitation forecasts of NUIST-CFS1.0 perform well in many regions of East Africa. In addition, positive skill of the deterministic and probabilistic forecasts is found over regions where the model shows positive performance in reproducing the impacts of the tropical SST anomalies on precipitation over East Africa.

Here, the prediction performance of NUIST-CFS1.0 is found to be on a level that is potentially useful but with plenty of room for future improvement, as is evident by the positive ROCSS over two-thirds of the study area. The GDS generally tends to lie between 0.6 and 0.8 over parts of Ethiopia, South Sudan, Uganda, and Kenya; this indicates the model’s potential use even for smallholder farmers. Therefore, a follow-up study will examine the added value of statistical downscaling techniques, including machine learning methods and statistical post-processing based on, for example, canonical correlation analysis (CCA), as well as explore their potential for applications at local scales over East Africa.

Acknowledgements.This work is supported by National Natural Science Foundation of China (Grant Nos. 42030605 and 42088101) and National Key R&D Program of China (Grant No.2020YFA0608004).

Electronic supplementary material:Supplementary material is available in the online version of this article at https://doi.org/10.1007/s00376-021-1180-y.