Prediction of seasonal precipitation can provide actionable information to guide management of various sectoral activities. For instance, it is often translated into hydrological forecasts for better water resources management. However, many studies assume homogeneity in precipitation across an entire study region, which may prove ineffective for operational and local-level decisions, particularly for locations with high spatial variability. This study proposes advancing local-level seasonal precipitation predictions by first conditioning on regional-level predictions, as defined through objective cluster analysis, for western Ethiopia. To our knowledge, this is the first study predicting seasonal precipitation at high resolution in this region, where lives and livelihoods are vulnerable to precipitation variability given the high reliance on rain-fed agriculture and limited water resources infrastructure. The combination of objective cluster analysis, spatially high-resolution prediction of seasonal precipitation, and a modeling structure spanning statistical and dynamical approaches makes clear advances in prediction skill and resolution, as compared with previous studies. The statistical model improves versus the non-clustered case or dynamical models for a number of specific clusters in northwestern Ethiopia, with clusters having regional average correlation and ranked probability skill score (RPSS) values of up to 0.5 and 33 %, respectively. The general skill (after bias correction) of the two best-performing dynamical models over the entire study region is superior to that of the statistical models, although the dynamical models issue predictions at a lower resolution and the raw predictions require bias correction to guarantee comparable skills.
Seasonal precipitation prediction can provide potentially actionable information to guide management of various sectoral activities. For instance, precipitation prediction is often translated into a hydrological forecast, which can be used to optimize reservoir operations, provide early flood or drought warning, inform waterway navigation, etc. As a primary input to soil moisture, precipitation prediction is also essential to agricultural management – farmers can take advantage of anticipated preferable climatic conditions or avoid unnecessary costs under expected undesirable conditions. Two types of models are commonly used for seasonal precipitation prediction: statistical and dynamical. Dynamical models, such as general circulation models (GCMs), include complex physical climate processes, while statistical models are purely data driven, relating observations and hydroclimate variables directly.
While both modeling approaches have produced skillful seasonal predictions for a variety of applications (e.g., Barrett, 1993; Hammer et al., 2000; Shukla et al., 2016), each has noteworthy drawbacks. Dynamical models often require a great amount of time to build and parameterize, whereas statistical models require considerably fewer resources (e.g., Mutai et al., 1998; Gissila et al., 2004; Block and Rajagopalan, 2007; Diro et al., 2008, 2011b; Block and Goddard, 2012). Dynamical models also suffer from their high sensitivity to initial uncertain conditions, particularly given a long lead time. Consequently, a number of simulations are typically produced, each with unique initial conditions, to provide a range of possible outcomes (e.g., Roeckner et al., 1996; Anderson et al., 2007). Furthermore, the outputs from dynamical models often require additional bias correction, typically using statistical methods, to better match observations (e.g., Ines and Hansen, 2006; Block et al., 2009; Teutschbein and Seibert, 2012). Statistical models, on the other hand, are highly dependent on substantial high-quality historical data to capture hydroclimatic patterns and signals, particularly extreme conditions, which is often not available. Additionally, statistical models are often linear by construction and may not well capture non-linear complex interactions and feedbacks. The physical nature of dynamical models, however, allows for prediction under non-stationary conditions and also when insufficient historical data are available, whereas statistical models, by construction, typically rely on stationary relationships (Schepen et al., 2012).
The spatial extent selected for statistical seasonal prediction is critical. It is not uncommon to simply assume homogeneity in precipitation across an entire study region; however, this limits addressing potential spatial variability. While this may be suitable for very broad regional planning, it is often ineffectual for operational and local-level decisions, particularly for locations with high spatial variability. This prompts the need for delineation of subregional-scale homogeneous regions, often defined through cluster analysis. Defining these homogeneous regions, however, is a non-trivial process. There are a variety of methods to delineate homogeneous regions, including comparing annual cycles (e.g., unimodal and bimodal distributions in precipitation) between stations (or grid cells), comparing station correlations with regional averages, applying empirical orthogonal functions (EOFs), various clustering techniques, and other methods of increasing complexity (e.g., Parthasarathy et al., 1993; Mason, 1998; Landman and Mason, 1999; Gissila et al., 2004; Diro et al., 2008, 2011b; Singh et al., 2012). In addition, delineation of the subregion size is also important to consider. Smaller-sized homogeneous subregions do not necessarily lead to improved predictions, as the noise at overly small scales can dominate any real signals representing spatial coherency of precipitation. For additional discussion regarding defining homogeneous subregions and cluster analysis, the reader is referred to Zhang et al. (2016) and Badr et al. (2015).
Spatial and temporal variability of June–September seasonal total
precipitation in western Ethiopia:
Precipitation in western Ethiopia peaks in the summer with approximately
70 % of annual total precipitation falling during the main raining season
– also known as the
The climate mechanisms affecting JJAS precipitation patterns in western Ethiopia are quite complex. Sea surface temperatures (SSTs) in the equatorial Pacific Ocean representing the well-known El Niño–Southern Oscillation (ENSO) phenomenon are considered a primary indicator of JJAS precipitation variability, with El Niño/La Niña often associated with deficit/excess of JJAS precipitation across the study region (e.g., NMSA, 1996; Camberlin, 1997; Bekele, 1997; Segele and Lamb, 2005; Diro et al., 2011a; Elagib and Elhag, 2011). Additionally, there is evidence of direct moisture transport from the Gulf of Guinea (equatorial Atlantic Ocean), the Indian Ocean, and the Mediterranean Sea, affecting Ethiopia's summertime precipitation (Viste and Sorteberg, 2013a, b). These moisture fluxes are often related to pressure patterns across the continent. For instance, the St. Helena High over the southern Atlantic Ocean or a high pressure over the Gulf of Guinea, coupled with a simultaneous low pressure over the Indian Ocean or a monsoon trough over the Arabian Peninsula, all bring intensified westerlies and southwesterlies that transport moist air across the Congo Basin to the western Ethiopian highlands (Segele et al., 2009; Williams et al., 2011). Similarly, the southwest Asian monsoon in the Indian Ocean, which has a strong positive relationship with concurrent JJAS precipitation in western Ethiopia, is associated with the Mascarene High over the southern Indian Ocean and a low pressure system near Bombay. During this monsoon season, the southeast trade winds in the Southern Hemisphere are channeled by the east African highlands while crossing the Equator and become a southwest monsoon flow. They are further diverted by the Turkana Channel, enhancing convergence with the westerlies/southwesterlies above the western Ethiopian highlands and bringing moisture to the region (Kinuthia, 1992; Nicholson, 1996, 2014; Camberlin, 1997; Slingo et al., 2005; Segele et al., 2009). In addition, the effect of other hydroclimate variables, such as Indian Ocean SST, local and regional atmospheric pressure systems (e.g., Azores High) also have notable influence on Ethiopia's precipitation variability (e.g., Kassahun, 1987; Tadesse, 1994; NMSA, 1996; Shanko and Camberlin, 1998; Goddard and Graham, 1999; Latif et al., 1999; Black et al., 2003; Segele and Lamb, 2005). Consequently, these large-scale climate variables may serve as potential predictors in statistical seasonal precipitation prediction models.
Ethiopia is vulnerable to fluctuations in precipitation given its reliance on rain-fed agriculture and limited water resources infrastructure. The majority of agriculture and infrastructure are in western Ethiopia, where water resources are relatively rich compared to other parts of the country (Awulachew et al., 2007). Operational precipitation predictions in Ethiopia have been issued by its National Meteorological Agency (NMA) since 1987 using an analog methodology (i.e., locating a similar climate condition in the past – an analog – to predict future conditions); however, this approach has produced only marginally skillful outcomes (Korecha and Sorteberg, 2013). For NMA's prediction, the country is divided into eight homogeneous regions for which NMA produces independent predictions. Similarly, others have also addressed seasonal prediction in Ethiopia contingent on both temporal and spatial precipitation patterns. Gissila et al. (2004) divide Ethiopia into four regions conditioned on the seasonal cycle and interannual variability coherence prior to prediction, while Diro et al. (2009) apply a similar approach but with dynamic cluster boundaries, allowing for different delineations for each rainy season. Segele et al. (2015) consider statistical precipitation predictions across Ethiopia as a whole, as well as for northeastern Ethiopia and at two Ethiopian cities. Block and Rajagopalan (2007) predict the average summertime (JJAS) precipitation over the upper Blue Nile Basin – a region they claim is homogenous at interannual timescales. Korecha and Barnston (2007) select an all-Ethiopia average precipitation index to characterize predictability broadly, with minimal attention to operational-level predictions. All of these studies focus on predicting regional average precipitation based on subjective clustering methods applying a limited number of stations or coarsely gridded data; no local predictions at a finer spatial scale are explored.
This study moves forward by exploring local-level seasonal precipitation prediction through the use of regional-level predictions, based on previous cluster analyses over western Ethiopia (Zhang et al., 2016). The advantages of defining homogeneous regions for seasonal prediction at operational (small) scales will be demonstrated by comparing approaches with and without undertaking a cluster analysis a priori. The combination of objective cluster analysis, spatially high-resolution prediction of seasonal precipitation, and a modeling structure spanning statistical and dynamical approaches makes clear advances compared to previous studies.
To evaluate high-resolution seasonal precipitation with versus without cluster analysis a priori, statistical models are developed and further compared with bias-corrected dynamical model predictions. Four scenarios are evaluated based on two criteria: (1) clustered vs. non-clustered and (2) direct vs. indirect. In the clustered case, predictions are produced for each homogeneous region (cluster) given a unique set of predictors. In the non-clustered case, the entire study region is considered as one cluster, and thus only one set of predictors is utilized for predictions. For the direct case, precipitation is predicted directly at the local level (grid scale); for the indirect case, the average precipitation within each homogeneous region is predicted first (as an intermediary) and then regressed to local-level (grid-scale) predictions. Combinations of the two criteria form four scenarios – clustered direct (C-D), non-clustered direct (NC-D), clustered indirect (C-I), and non-clustered indirect (NC-I) predictions.
Using a
Regionalization map of eight homogeneous regions marked by different colors, with country boundaries and river profiles. The figure is based on Zhang et al. (2016).
Many studies have investigated statistical models for seasonal climate
prediction. These studies vary by preclassification of predictor or
predictand regime, predictor selection process, and statistical methods. For
example, Hertig and Jacobeit (2011) investigate SST
regimes as potential predictors for subsequent precipitation and temperature
in the Mediterranean region. Through techniques including multiple
applications of principal component analysis (PCA), 17 stationary SST regimes were identified. Gerlitz et
al. (2016) apply a
Season-ahead (March–May) or month-ahead (May) large-scale climate variables
that are physically relevant in potentially modulating moisture transport to
the basin (or cluster) are selected as potential predictors. Four climate
variables are selected here for further evaluation based on outcomes of the
aforementioned prediction studies: SST, sea-level pressure (SLP), geopotential height (GH) at
500 mb, and surface air temperature (SAT). All climate variables are from
the National Centers for Environmental Prediction and National Center for
Atmospheric Research (NCEP/NCAR) reanalysis dataset (Kalnay et al., 1996) at
a 2.5
Those potential predictors are first transformed through PCA (Jolliffe, 2002). PCA is a common approach in climate modeling to reduce the dimensionality of predictors and remove multi-collinearity, while simultaneously extracting the most dominant signals from the potential predictors, typically reflected in the first few principal components (PCs). Since PCA is independent of the predictand, retaining the first few PCs as predictors, in lieu of the original variables, also helps to reduce artificial prediction skill.
Flow chart of data processing for predictors into the statistical model. Numbers framed by dashed lines correspond to the procedures listed in the context. Note: pre. – precipitation, t-s – time series, avg. – average.
Subsequently, a certain number of PCs are used as the direct inputs into a MLR model, otherwise known as the principal component regression (PCR). PCR is performed in a “drop-one-year” cross-validation mode to reduce overfitting effects and therefore avoid overestimation of prediction skill. This requires reconstructing the principal components for the dropped year and then multiplying the coefficient estimates with each reconstructed PC, respectively, in order to obtain the final predicted value for the dropped year (e.g., Block and Rajagopalan, 2009; Wilks, 2011). A detailed methodology is provided below.
Justifiable climate regions globally for selecting predictors.
Correlation map between mean JJAS seasonal precipitation time series in Cluster 5 and global SST under cross-validation, with correlations lower than the 99 % significance level masked out (one-tail test).
Equations of linear regression panel models under four scenarios.
To avoid overfitting, the entire process including predictor selection and
statistical modeling is processed using cross-validation. To start,
drop-one-year precipitation observations for JJAS averaged across the region
and each cluster are spatially correlated independently with each global
climate variable. As a result, there are total of 1044 global correlation
maps given the 29-year time series, eight clusters plus one non-cluster, and
four climate variables. Hence, a program to automatically select highly
correlated and justifiable regions as predictors is developed. The following
steps describe the subsequent statistical modeling process (Fig. 3):
Grid cells within each justifiable region (e.g., equatorial Pacific; Fig. 4)
with correlation above the 99 % significance level are identified
(Fig. 5). For regions containing grid cells with both positive and negative
correlations, the number of the identified grid cells in each sign is
counted. If a greater number of grid cells is associated with significant
positive correlation, for example, only grid cells with positive correlations
are kept for the following steps, and vice versa. The top 10 % of the identified grid cells with the highest correlation
in each region are then selected in order to boost the potential model skill. For each region, data of the selected grid cells within the region are
spatially averaged (defined as “prepredictors”). Prepredictors are standardized, combined, and transformed through PCA
for each cluster or non-cluster and each dropped-year analysis separately. The top PCs from the PCA with a total of
95% variance explained are used as predictors in PCR. For the
direct case, PCR is used to directly predict the grid-level
precipitation; for the indirect case, PCR is used to predict the
intermediate cluster-level precipitation. For the indirect case only, cluster-level predictions are
regressed to the grid level. Note that the downscaling of cluster-level
predictions to grid-level predictions is also cross-validated to avoid
overfitting.
For the four scenarios, the model structures are quite similar but have
subtle differences which could lead to evidently different outcomes
(Table 1). Under the NC-D (Eq. 1a, b) and C-D scenarios (Eq. 2a, b), the
time series of JJAS seasonal total precipitation in each grid cell (i.e., at
local level) is used as the direct predictand (
Cluster-level predictions and observations under the C-I and NC-I scenarios, with drop-one-year cross-validation. The 95 % envelope shows the 95 % confidence interval constructed using model errors.
Pearson correlations between grid-level observations and predictions under the four scenarios, with the clustering boundary delineated roughly in black.
The North American Multi-Model Ensemble (NMME; Kirtman et al., 2014) is an
experimental multi-model system consisting of coupled dynamical models from
various modeling centers in North America. To our knowledge, it is also the
most extensive multi-model seasonal prediction archive. The NMME provides
gridded climate predictions that cover regions globally and with different
lead times. The hindcasts of monthly mean precipitation are easily
accessible through the International Research Institute for Climate and
Society (IRI) website
(
Grid-level RPSS (%) under the four scenarios using climate variables as predictors, with the clustering boundary delineated roughly in black.
The NMME predictions for each of the 10 models are bias corrected by applying
probability mapping (e.g., Block et al., 2009; Teutschbein and Seibert, 2012;
Chen et al., 2013) under cross-validation, subject to the observational
dataset from NMA (Fig. 6). This is performed on a grid-cell-by-grid-cell
basis on standardized data (the NMME dataset is reshaped to
0.1 Fit gamma distributions to drop-one-year time series from each
observed and NMME grid cell; for NMME, this is performed on an individual
model basis using all ensemble members available. (Goodness-of-fit tests
indicate gamma distributions are appropriate; results not shown.) Translate gamma distributions into cumulative distribution functions
(CDFs). For any given dynamical model prediction at the grid-cell level, a
corrected prediction value is attained by mapping from the modeled CDF to the
observed CDF and applying the inverse gamma distribution. This is repeated
for all grid cells, all NMME models, and all dropped years.
After correction, the gamma CDFs of predictions and observations approximately
match (Fig. 6a). Additionally, each ensemble still retains its variability
over time, though the overall ensemble mean is shifted to closely match
observation (Fig. 6b).
Pearson correlations are used to measure the standardized covariance between observations and predictions. Ranked probability skill scores (RPSSs; Wilks, 2011) are also evaluated to determine categorical skill based on probabilistic predictions. Here, the data are split into three equal terciles representing below-normal, near-normal, and above-normal conditions. A perfect prediction yields an RPSS of 100 %, and a prediction with less skill than climatology (long-term averages) yields an RPSS of less than zero. Median RPSS values from all 29 years are reported.
Box plots of grid-level RPSS (%) for 10 dynamical models from
NMME
Pearson correlations between grid-level observations and ensemble mean of bias-corrected predictions for 10 dynamical models from NMME, labeled with the same number as listed in the context.
Correlation coefficients (“Corr.”) and RPSS for (drop-one-year cross-validated) predictions at cluster level compared to observations under the C-I and NC-I scenarios.
Correlations between cluster-level model predictions and observations range
from
Grid-level Pearson correlation and RPSS statistics.
At the grid scale, depending on the case (direct or indirect), and for different clusters, correlations between predictions and observations can favor the clustered case or the non-clustered case (Fig. 8). In general, the indirect model provides a smoother pattern of correlations, with grid cells showing a negative correlation in the direct case now improved to near or above zero (Fig. 8). For example, Cluster 5 under the indirect case illustrates a more consistent positive correlation within the cluster. Some parts of the study region reach a correlation over 0.6, such as central-northwestern Ethiopia (Cluster 5), which is consistent with the region of high cluster-level prediction skill. The percentage of grid cells with correlations passing the 95 % significance test is the highest for the NC-D case (Table 3); however, some locations demonstrate the lowest skills among all four scenarios.
Similar findings are evident by evaluating the RPSS, except for Cluster 8; instead of improving with increased RPSS in the indirect case, the grid-scale predictions deteriorate given poor cluster-level prediction (for the C-I case). However, the percentage of grid cells with positive RPSS values overall for the C-I case is still the second highest after the NC-I case (Table 3), indicating the indirect cases are superior in terms of the number of grid cells with improved skill compared to using climatology, particularly for grid cells associated with skillful intermediate cluster-level predictions. The predictions are most skillful for the same region of central-northwestern Ethiopia (Cluster 5; Fig. 9) with 87 % of its grid cells showing positive RPSS and a spatial-average RPSS value of 15 % under the C-I scenario (Table 4).
The RPSS values based on the prediction ensembles of each dynamical model
improve remarkably after bias correction. The median RPSS values over all the
grid cells are now close to zero (Fig. 10) with two models, NASA-GMAO and
NCEP-CFSv2, showing the highest RPSS value (
Grid-level Pearson correlation and RPSS statistics for grid cells within Cluster 5.
Within a certain cluster, statistical models may perform better than all
dynamical models. For example, for Cluster 5, all statistical models show
higher average RPSS values than those of all dynamical models (Table 4). The
percentage of grid cells with significant correlation reaches 61 % for
the statistical model under the NC-D scenario, compared to the highest value of
47 % among all the dynamical models. Similarly, the percentage with
positive RPSS achieves 87 % under the C-I scenario as opposed to 66 % for
dynamical models. Note that the dynamical models also produce raw predictions
in a lower spatial resolution (1
This study demonstrates the potential for applying season-ahead large-scale climate information to predict high-resolution precipitation using a statistical modeling approach. Skillful and credible predictions are produced for some regions in western Ethiopia, particularly under a clustered indirect statistical approach. At the regional scale, the approach shows promise for northwestern Ethiopia (Clusters 1, 3, 5, and 7), particularly compared to current NMA operational forecasts, which are only moderately more skillful than climatology (Korecha and Sorteberg, 2013). The regional average RPSS in this study under the clustered case ranges from 10 to 33 % for northwestern Ethiopia, as opposed to values under 6 % for NMA operational forecast (Korecha and Sorteberg, 2013). The approach adopted here also advances on previous studies (Gissila et al., 2004; Block and Rajagopalan, 2007; Korecha and Barnston, 2007; Diro et al., 2011b; Segele et al., 2015) by first applying an objective cluster analysis and then conditionally constructing high-resolution predictions. A unique set of predictors is applied to each cluster, which contributes to superior prediction performance at cluster levels in northwestern Ethiopia, as compared with predictions from the non-clustered approach. Grid-level prediction under the clustered indirect case also reduces the effect of overfitting relative to the direct case and improves negative RPSS values to near or above zero; that said, the non-clustered direct case also illustrates higher correlation and RPSS values on average.
A total of 2 out of 10 NMME dynamical models, NASA-GMAO and NCEP-CFSv2, demonstrate overall superior performance to the statistical models; however, for certain regions such as Cluster 5, the performance of statistical models under the clustered indirect and non-clustered direct cases is still superior. It is also worth noting that the statistical model predictions are at a 100 times finer spatial resolution than the dynamical models, providing additional advantages at the local scale, when skillful. Nevertheless, improvements in dynamical models continue and their application to seasonal precipitation prediction is likely to grow (e.g., Palmer et al., 2004; Saha et al., 2006; Lim et al., 2009).
Relatively poor prediction performance is evident in some locations such as southwestern Ethiopia and regions outside Ethiopia, where the hydroclimatic processes that produce precipitation might be driven by local factors or other regional climate patterns rather than large-scale climate variables identified in this study. A previous study (Zhang et al., 2016) has shown that the influence of ENSO on JJAS precipitation in western Ethiopia decreases generally from north to south, and is likely one of the reasons why skills are relatively low in southwestern Ethiopia. Cluster 5 was also identified with the strongest connection to equatorial Pacific SST (Zhang et al., 2016), which is consistent with the highest skill found in this study. Other regions with low prediction skill show relatively strong connections to SST in neighboring oceanic regions. However, connections with those climate patterns appear to be less robust than with ENSO, making the predictions in their associated regions less skillful. This is also consistent with the findings from other studies that even though all three oceans (Indian, Atlantic, and Pacific) affect the JJAS precipitation in western Ethiopia, the Pacific Ocean still plays the greatest role (Segele et al., 2009; Omondi et al., 2013).
The southwest Asian monsoon over the Indian Ocean may also be critical in determining the precipitation, given that the clusters with better prediction skills lie along the pathway of the monsoon. Based on the global concurrent correlation maps between JJAS precipitation and SLP for each cluster, Clusters 5 and 7 – the two clusters with the best skills – are the only ones that are strongly and negatively correlated with SLP near Bombay, and in the meantime strongly and positively correlated with the SLP at the eastern equatorial Pacific Ocean. The former indicates that a strong southwest Asian monsoon is associated with higher JJAS precipitation amount, and vice versa. The latter indicates that a high surface pressure over the eastern equatorial Pacific Ocean often accompanied by cold SST and a raining pattern – a La Niña phenomenon – also brings higher JJAS precipitation to western Ethiopia, and vice versa. Cluster 2 – one of the worst predicted clusters – shows moderately strong negative correlation with SLP near Bombay; however, it is also correlated strongly and negatively with SLP in the southern Indian Ocean (a high pressure system that drives the monsoon toward the low pressure system near Bombay), indicating that high JJAS precipitation in Cluster 2 is not necessarily associated with a strong southwest Asian monsoon. Moreover, its correlation with SLP over the equatorial Pacific Ocean is nonsignificant. Considering, in general, El Niño suppresses the monsoon and La Niña increases it (Kumar et al., 2006), strong correlations with both ENSO and the monsoon in the correct direction, such as for Clusters 5 and 7, indicate a double insurance over their association with the southwest Asian monsoon. Therefore, clusters which are more affected by the southwest Asian monsoon over the Indian Ocean, particularly coupled with the influence of ENSO, are likely to be more promising in their prediction skills.
Additional prediction features also warrant future attention, including longer prediction lead times and evaluation of other relevant characteristics (e.g., intraseasonal dry spells, seasonal onset or cessation). As observational datasets continue to grow, data-driven cluster analyses and statistical modeling approaches may be expected to improve. Careful analysis of possible significant trends in the data is also warranted; a region with a relatively high correlation may be selected solely based on trends in predictors and observations. For shorter time series, such as the data used in this study, trend analysis may not be reliable; detrending can also reduce evidence of large-scale decadal climate signals. Improving predictive capabilities may not be a complete panacea, but it can continue to be an important part of a decision maker's portfolio as they cope with hydroclimatic variability and its inherent risks.
The National Centers for Environmental Prediction and
National Center for Atmospheric Research (NCEP/NCAR) reanalysis dataset can
be accessed through the National Oceanic and Atmospheric Administration
(NOAA) Earth System Research Laboratory (ESRL) website
(
The NMME hindcasts are available through the International Research Institute
for Climate and Society (IRI) website
(
The gridded precipitation dataset in western Ethiopia is available upon
request from NMA (
The authors declare that they have no conflict of interest.
This article is part of the special issue “Sub-seasonal to seasonal hydrological forecasting”. It does not belong to a conference.
This study was supported by NASA project NNX14AD30G and NSF PIRE project 1545874. We acknowledge the National Meteorological Agency of Ethiopia for sharing data. We also want to thank the reviewers for their suggestions in improving this work. Edited by: Quan J. Wang Reviewed by: Yared Bayissa and Lars Gerlitz