Identification of deficiencies in seasonal rainfall simulated by CMIP5 climate models

An objective technique for analysing seasonality, in terms of regime, progression and timing of the wet seasons, is applied in the evaluation of CMIP5 simulations across continental Africa. Atmosphere-only and coupled integrations capture the gross observed patterns of seasonal progression and give mean onset/cessation dates within 18 days of the observational dates for 11 of the 13 regions considered. Accurate representation of seasonality over central-southern Africa and West Africa (excluding the southern coastline) adds credence for future projected changes in seasonality here. However, coupled simulations exhibit timing biases over the Horn of Africa, with the long rains 20 days late on average. Although both sets of simulations detect biannual rainfall seasonal cycles for East and Central Africa, coupled simulations fail to capture the biannual regime over the southern West African coastline. This is linked with errors in the Gulf of Guinea sea surface temperature (SST) and deficient representation of the SST/rainfall relationship.


Introduction
The timing and seasonality of precipitation is of critical importance to the many African stakeholders who depend upon the seasonal rains for agricultural and domestic purposes. Failure or delays in these rains can lead to significant socio-economic impacts (Vizy et al 2015). Future changes in climate will be felt not only through changes in mean precipitation, but also through altered seasonality, which in turn influences the growing season and crop yields (Vizy et al 2015), the length of the malaria transmission season (Tanser et al 2003), the supply of hydroelectric power (Yamba et al 2011, van Vilet et al 2016 and surface water supplies (de Wit and Stankiewicz 2006). Producing reliable projections of the impact of climate change in a range of sectors therefore requires an accurate representation of seasonality within the climate projections being utilized.
Seasonality is sensitive to changes in atmospheric circulation patterns (Shongwe et al 2009, Lee andWang 2014) and diagnosing and interpreting such changes requires a robust understanding of the dynamics and drivers of the seasonal cycle. In order to use global climate models (GCMs) to investigate the physical mechanisms driving the seasonal cycle of precipitation, and assess the reliability of future impact projections, it is necessary that GCMs are able to represent the seasonality of African precipitation.
In this study we take a continental scale approach to the assessment of the seasonality of African precipitation in the Atmospheric Model Intercomparison Project (AMIP) and historical experiments as part of the Coupled Model Intercomparison Project Phase 5 (CMIP5) simulations (Taylor et al 2012). An objective technique is applied using cumulative rainfall anomaly to calculate the onset and cessation of wet seasons (Liebmann et al 2012, Dunning et al 2016. This technique enables examination of the nature of the seasonal cycle, in terms of the number of wet seasons experienced per year (seasonal regime), the patterns of rainfall advance and retreat, and the timing of the wet seasons. This paper aims to address three main questions: • Can models realistically represent contrasting annual and biannual seasonal precipitation regimes across the continent? • How well is the seasonal progression of rainfall represented, including spatial patterns of rainfall advance and retreat throughout the year? • How well is the timing of the rainy seasons captured?
Many previous studies examining the representation of African precipitation in CMIP models have assessed mean rainfall amount and interannual variability for fixed seasons (Kumar et al 2014, Lee and Wang 2014, Mehran et al 2014, Maidment et al 2015. Those studies that do consider the seasonal cycle of precipitation either take averages over large areas, smoothing out much of the variability, and neglecting the seasonal progression of the rains (Dike et al 2015) or use metrics such as seasonality skill score (Koutroulis et al 2016), sum of squared errors (Yamana et al 2016), or space-time root-mean-square error (Flato et al 2013) which primarily focus on long-term monthly data. The methodological approach used here enables comparison between models, notwithstanding timing discrepancies. Additionally, the proven agricultural relevance of this onset/cessation metric means this evaluation of climate models is informative for impact studies (Dunning et al 2016).
Studies that examine the onset and cessation of rainfall in climate models, taking into account the seasonal progression of rains, have previously focused on specific regions such as East Africa . The continental approach taken here complements these regional studies, providing insight into the links between timing biases in different regions and the links with largescale progression of the wet season across regions. Several of the previous studies into seasonality have used regional climate models, such as the CORDEX regional models (West Africa: Mounkaila et al 2015, Southern Africa: Shongwe et al 2015, East Africa: Endris et al 2013. While these models are valuable in providing regional detail, they are driven at their lateral boundaries by GCMs, and therefore understanding the seasonality in GCMs remains of key importance. Furthermore, some impact projections use GCMs, and few conduct thorough assessments of the representation of precipitation seasonality, potentially using models with significant deficiencies in seasonality representation (Yamba et al 2011, Hamududu and Killingtveit 2012, Caminade et al 2014, van Vilet et al 2016. The United Nations Development Programme (UNDP) Climate Change Country Profiles contain multi-model mean projections of annual and seasonal temperature and precipitation for a range of developing countries and were designed to facilitate climate change assessment with minimal computational expense (McSweeney et al 2010), and are currently utilised in both agricultural (Adhikari et al 2015) and ecological (Laloë et al 2014) sectors. However, the same set of 15 GCMs are used for all countries, and hence multimodel mean projections in the locally-defined seasonal means may be inaccurate if the models used contain seasonal cycle timing biases.
This study presents, for the first time, a continentwide evaluation of climate model representation of African rainfall seasonality. It aims to better characterise the representation of the seasonal cycle in terms of regime, progression and timing, in order to provide more detail on model deficiencies and their causes, and to facilitate better application of such models for climate change impact assessments.

Methods and data
In order to assess rainfall seasonality in terms of seasonal regime (annual or biannual), progression and timing, we apply the method of Dunning et al (2016) which extends the methodology of Liebmann et al (2012). Firstly, at each grid point the ratio between the amplitude of the first and second harmonics is computed to determine whether an annual (one wet season/year) or biannual (two wet seasons/year) regime dominates. Secondly, the period of the year when the wet season occurs is determined by identifying the minima and maxima in the climatological cumulative daily mean rainfall anomaly, to account for seasons that span multiple calendar years. Dunning et al (2016) introduces a method for doing this for biannual regimes, accompanying the method for annual regimes in Liebmann et al (2012). Following the identification of the climatological seasons, onset and cessation dates are calculated for each season and year individually by identifying the minima and maxima in the daily cumulative rainfall anomaly over that season. The climatological cumulative daily mean rainfall anomaly and daily cumulative rainfall anomaly are computed independently for each climate model and dataset. Onset and cessation dates are not calculated for the first or last years of each dataset. The approach was found to be applicable for datasets with contrasting rainfall biases, producing contemporaneous onset/cessation dates (Dunning et al 2016), and thus can be applied to model simulations without the need for bias correction. The method also permits seasons to span different time periods thus enabling the examination of timing differences and comparison of wet seasons that do not coincide exactly across models and observations. By comparison with agricultural onset methods proposed by Marteau et al ( For calculation of onset and cessation dates, daily precipitation data is required. Forty-six models were chosen from the CMIP5 generation of models (Taylor et al 2012), based on the availability of daily precipitation for an AMIP simulation (which applies observed sea surface temperature (SST) and sea ice and realistic radiative forcings) or CMIP5 historical simulation (which includes a fully coupled ocean and is driven by historical radiative forcings) from the British Atmospheric Data Centre (BADC). 28 AMIP simulations were used over 1979-2008 and 39 CMIP historical simulations were used over 1979-2005 (some variations in dates; supplementary table S1, available at stacks.iop.org/ERL/12/114001/mmedia). Models, name of institute and horizontal resolution are listed in table S1. Only the first ensemble members (r1i1p1) are considered since differences across ensemble members were found to be minimal, and much smaller than the inter-model differences.
To allow construction of multi-model means, the data were re-gridded using bilinear interpolation to the GPCP 1DD 1 • × 1 • grid. For some models this is a large increase in resolution; results were compared using a 2 • × 2 • grid for the CMIP historical simulations and found to be unchanged. Following re-gridding, onset and cessation dates were calculated. For the individual model results in the supplementary information, the method is applied at native resolution (table S1).
In order to compare rainfall over the southern West African coastline and SSTs over the Gulf of Guinea (section 4), monthly SST data was obtained for 37 of the 39 historical simulations from BADC. Monthly HadISST observed SST and monthly AMIP forcing SST data were obtained at 1 • × 1 • horizontal resolution.
A reference dataset was required for comparison, to facilitate the assessment of the ability of CMIP5 model simulations to represent seasonality over Africa. Inaccuracies in ERA-Interim reanalysis precipitation data have been identified ( For analysis of timing, Africa was divided into a number of regions (figure S1). These were chosen to primarily exhibit an annual or biannual regime and contain broadly similar onset and cessation dates. The area north of 15 • N was not included in the analysis due to the dry climate. The GPCP annual/biannual categorisation was used as a basis for the region mask; model simulation grid-points were considered as belonging to a region if they fell within the correct geographical area and had the correct annual/biannual categorisation. Hence figure S1 shows the maximum possible extent of the regions; for some models fewer points were included if they had an incorrect categorisation. Figure 1 illustrates the application of the Dunning et al (2016) method over the Horn of Africa and the Sahel (region map: figure S1/4(b)). Average rainfall rate in each of the four/two seasons is calculated for the long term climatology (all available years; coupled: 1980-2004, atmosphere-only: 1980-2007, observations: 1982-2013) for standard meteorological seasons (Horn of Africa: short rains October-December, long rains March-May; Sahel: wet season July-September, dry season October-June; hashed bars) and when the seasons are determined dynamically for each model and year (solid bars) using the method of Dunning et al (2016) which allows seasons to shift in space and time from year to year and across different models. Rainy seasons are differentiated more clearly from the dry seasons when the seasons are determined dynamically, as the method accounts for interannual variability in seasonal timing and model timing biases. This better seasonal distinction is obtained for all regions (figures S2-S3) with wet/dry season difference in average rainfall rate 33% larger when seasons are defined dynamically (annual regions). This demonstrates the advantage of examining seasonality using a method that objectively identifies wet seasons in models, and would enable the analysis of future changes in seasonality, while accounting for errors in the mean state. The socioeconomic importance of the East African long rains, necessitates extensive climate model analysis of this season, where use of a method that better distinguishes the long rains (as seen in figure 1

Results
First the representation of annual and biannual seasonal regimes in the model simulations is examined.   2(a)). The atmosphere-only simulations capture the biannual regime over these regions ( figure 2(b)), and while the area experiencing a biannual regime is larger than in the observations, there is good agreement between figures 2(a) and (b). While the coupled simulations capture the biannual regime for the Horn of Africa and the equatorial strip, the biannual regime over SWAC is not represented ( figure 2(c)). The lack of such a season over Nigeria in HadGEM2-ES has also been identified by Dike et al (2015), but causes were not attributed.
In both experiments an annual regime is found over southern Africa, West Africa and the Sahel, as in the observations.
Secondly, the representation of the spatial patterns of seasonal progression of precipitation is examined. Figure 3 shows the multi-model mean onset and cessation dates for both annual and biannual regimes.   The third aspect of seasonality examined was seasonal timing (figure 4), full results for all models and regions (figure S1) are in figures S24-S26. The multimodel mean onset and cessation dates for five of the regions (figure 4(a)) show the models generally simulate the wet season at the correct time of year. Notable differences include early onset over the Sahel (23 ± 15 days on average for atmosphere-only, 25 ± 18 days for coupled simulations, where the error relates to standard deviation across models), longer short rains over the Horn of Africa (cessation ∼16 ± 11 days later) and later long rains (onset/cessation ∼19 ± 13 days late for coupled simulations) over the same region. The late bias of the long rains in coupled simulations, with cessation in early June (on average) may explain some of the dry bias previously found in the long rains over East Africa in CMIP5 models (Yang et al 2015a). The overestimation of the short rains, typically found in such models (Yang et al 2015a), is replicated here in the combination of longer season length, of 14 ± 14 days longer for AMIP and 25 ± 15 days longer for CMIP historical simulations (figure 4(a)), and similar average rainfall per day ( figure 1(a)).
Most coupled simulations do not capture the biannual regime over SWAC (figure 2(c)), hence the mean onset and cessation dates for the annual regime over this region are shown in figure 4(a), while the season over SWAC is longer than is found over the west Sahel, it is shorter than atmosphere-only and observations for SWAC, with onset occurring during the first rains, and cessation occurring during the second rains. Biannual onset and cessation dates for one of the two coupled simulations that simulate a biannual regime here (MIROC4h and BCC-CSM1-1-M) show good agreement with observations, but the onset dates are late in BCC-CSM1-1-M.
Mean onset and cessation dates were computed for each simulation for 13 regions with an annual regime (see figure S1) and the Horn of Africa, and compared with mean observational onset and cessation dates. For atmosphere-only simulations the mean difference from observations for onset ranged between −18 ± 14 days to +6 ± 10 days (negative values denote early onset, positive is late), with the onset over the Sahel exhibiting larger differences of −23 ± 15 days. Cessation dates exhibited smaller mean differences, with differences ranging between −5 ± 10 days to 8 ± 15 days for the different regions. For coupled simulations the mean difference from observations for onset ranged from −15 ± 10 days to +3 ± 13 days, with larger differences over the Sahel (−25 ± 18 days) and over the west coast of southern Africa. The discrepancy over this southern region is likely to be related to overestimation of the area that experiences a winter rainfall regime shown in figure 3, and exclusion of dry areas associated with the Namibian Desert (Liebmann et al 2012, Engelbrecht et al 2015. For cessation, differences from the observations are of similar magnitude, with differences ranging from −12 ± 10 days to +15 ± 11 days, with the west coast of southern Africa again an outlier, showing a much higher value. Individual results for each model and region are included in the supplementary information. Overall, figures 2-4 (and figures S4-S23) indicate that the model simulations realistically capture the seasonal progression of the rainy seasons over Africa.
However, coupled simulations overestimate the areal extent of the winter rainfall regime in South Africa, leading to large timing errors over South Africa (40-50 days on average, region 15 in figures S24-S25) and importantly fail to capture the biannual regime over SWAC. The potential for detrimental impacts of the intervening dry season on agricultural yields (Odekunle, 2007)

Simulation of the Little Dry Season in southern coastal West Africa
The biannual regime over SWAC comprises four seasons; a dry season from November-March, the first wet season from April-June and the second wet season from mid-September-October. These two wet seasons are separated by a break in the monsoon rains during July-August, known as the 'August Break' (Chineke et al 2010) or 'Little Dry Season' (LDS) (Odekunle 2010). The LDS can be a useful period for weeding and spraying of crops; when the LDS is too early, long or intense, yields can be reduced (Adejuwon and Odekunle, 2006). Both the mean annual cycle of precipitation in observations (blue line) and atmosphere-only (red line), over the LDS region ( figure 5(b)), show the biannual seasonal cycle. However, figure 5(b) confirms that coupled simulations (green line) do not capture the correct seasonal cycle.
The misrepresentation of the Little Dry Season is related to the southward Intertropical Convergence Zone (ITCZ) bias in coupled simulations (Richter and Xie 2008, Roehrig et al 2013, Toniazzo and Woolnough 2014, Monerie et al 2017. This bias has been connected with the presence of warm Atlantic SST biases in the Gulf of Guinea (Vizy and Cook 2001, Cook and Vizy 2006, Roehrig et al 2013, via influences on the meridional temperature gradient. Furthermore, the strength of the West African Monsoon is sensitive to SSTs in the equatorial cold tongue (Patricola et al 2012), such that sporadic warming in this region resulting from Atlantic Niños (Nnamchi et al 2015) weakens the monsoon circulation and monsoon rains are confined to the Guinea Coast (Chang et al 2008). Hence, the failure of the coupled simulations to reproduce the eastern equatorial cold tongue in boreal summer (Richter andXie 2008, Patricola et al 2012) may be associated with the restricted northward progression of the ITCZ. More locally, the reduction in rainfall during the LDS is accompanied by cool SSTs in the northern Gulf of Guinea, in particular, between 8 • W-2 • E, and 3 • N to the West African coastline ( figure 5(a), pink box), resulting from ocean upwelling (Odekunle and Eludoyin 2008) (figure 5). Adejuwon and Odekunle (2006) proposed that the LDS was a consequence of these local cooler SSTs leading to the inter-tropical discontinuity travelling further inland, while Odekunle and Eludoyin (2008) suggested that cool SSTs increased static stability over this region, inhibiting convection. The correct representation of the LDS in atmosphere-only simulations with prescribed SSTs and incorrect representation in fully coupled simulations ( figure 5(b)), combined with previous research (Adejuwon and Odekunle 2006, Odekunle andEludoyin 2008), suggests an SST driver for the LDS. Thus the relationship between SSTs in the northern Gulf of Guinea region (8 • W-2 • E, and 3 • N to the West African coastline; pink box in figure 5(a)), and LDS rainfall is explored to determine if this could explain the misrepresentation (figure 5). Observational datasets were used to determine the mean period of the LDS (2 July-17 September), and the region influenced by this regime (SWAC; blue dots in figure 5(a)). Linear regression was used to examine the relationship between LDS rainfall and July-August-September (JAS) SST (figure 5(c)).
Observations and atmosphere-only simulations exhibit the expected relationship with colder SST anomaly years resulting in lower LDS rainfall, yet coupled simulations fail to represent this association ( figure 5(c)).
In addition, none of the coupled simulations realistically represent the SST seasonal cycle over the region of interest identified by Adejuwon and Odekunle (2006), the minima are in September rather than August, and the amplitude of the SST seasonal cycle is only ∼55% of that in HadISST, at 2.3 K compared with 4.3 K in HadISST and 4.1 K in AMIP. The coupled historical simulations for MIROC4h and BCC-CSM1-1-M contain an LDS at a few coastal grid-points, evident in most ensemble members, and also exhibit SST anomalies in the lowest 10% for July and August, compared with all coupled simulations (brown/pale pink dashed line, figure 5(d)).
These findings are consistent with the lack of a LDS in coupled simulations being related to unrealistic representation of the SST seasonal cycle, and an incorrect interannual relationship between SST and rainfall. The seasonal cycle of SST is an important driver of seasonality in rainfall, with previous studies identifying a significant influence of the decline in equatorial SST from April to July on the development of the monsoon (Okumura and Xie 2004) and a delay in the phase of the Atlantic SST seasonal cycle resulting in an increase in late rainy season precipitation (Monerie et al 2017). It may be inferred therefore that models that represent SST seasonality well are more likely than others to capture the seasonality of rainfall accurately. However, a good representation of SST is not the only factor in determining models' ability to capture rainfall. When analyzing the Atlantic ITCZ structure, Siongco et al (2015) found that model horizontal resolution had a large influence on the marine ITCZ. MIROC4h and BCC-CSM1-1-M have higher spatial resolution than many of the other models used and better representation of the SST seasonal cycle, with lower SST anomalies in JAS. Yet, it would be over-simplistic to argue that resolution and accurate representation of SST alone determine models' ability to capture the LDS at a few grid points. Monerie et al (2017) noted that the influence of SST varies across climate models, which may explain why some models with similar SST seasonal cycle and resolution (e.g. EC-Earth) do not contain an LDS. Of the 18 models analysed by Cook and Vizy (2006) only four correctly captured the relationship between Gulf of Guinea SST anomalies and West African Monsoon precipitation anomalies; hence even with the correct SST the coupled simulations may not produce a LDS due to inaccurate representation of the variability. Furthermore, the monsoon acts to cool SST (Okumura andXie 2004, Hagos andCook 2009), hence correct representation of SST seasonality is both a cause of and response to the accuracy of precipitation seasonality. The preceding discussion thus highlights two likely causes of model bias in the representation of the West African Monsoon and provokes a number of additional questions. Further detailed analysis of such mechanisms is merited, along with other LDS drivers, including the deflection of south-westerly winds to westerly winds (Olaniran 1989), which could be addressed through idealized model integrations and further analysis of observations.

Conclusions
The representation of precipitation seasonality in atmosphere-only and coupled historical CMIP5 climate model simulations has been evaluated across Africa using observations and an objective methodology for quantifying seasonal characteristics (Liebmann et al 2012, Dunning et al 2016. Using this methodological approach demonstrates the presence of a biannual regime over regions with timing biases, where the use of standard meteorological seasons may suppress the perceived seasonal cycle, and thus can be used to compare model simulations, notwithstanding timing biases. Overall, the CMIP5 simulations capture the gross seasonal cycle of African precipitation on a continental scale. The patterns of seasonal progression of the rainy season are well-represented, and the atmosphereonly simulations realistically simulate biannual rainfall regimes for the Horn of Africa, Equatorial Africa and a region on the southern West Africa coastline (SWAC). Patterns of seasonal progression over southern Africa are the consequence of interactions between the ITCZ, the South Atlantic Anticyclone, mid latitude westerlies and the Angola Low, where the tropical-extratropical cloud bands that bring summer rainfall across southern Africa form (Reason et al 2006). The large-scale similarity in spatial rainfall progression patterns over Southern Africa across observations, atmosphere-only and coupled integrations suggest that the models capture these interactions which lead to the non-zonal patterns observed in this region (Reason et al 2006, Shongwe et al 2009, Shongwe et al 2015. Seasonal timing in atmosphere-only simulations demonstrates good agreement with observations (most mean onset/cessation dates agree within 18 days) although both sets of simulations exhibit an early onset of the wet season over the Sahel (∼25 days early). Additionally, coupled simulations exhibit timing biases over the Horn of Africa (long rains ∼20 days late).
Although the seasonal cycle is generally represented well in atmosphere-only model runs, in regions influenced by annually changing SST, SST biases in coupled simulations lead to errors in rainfall patterns. The incorrect annual seasonal regime found over SWAC (biannual in observations) has been linked to an unrealistic seasonal cycle of SST and erroneous SST/rainfall relationships in this region. Additional work is merited to further elucidate exact mechanisms that determine the realistic representation of certain features and improve physical mechanisms represented by coupled simulations. Climate model projections should be treated with caution, and projections of future agriculture production in this region are likely to be unrealistic, due to the significant impact of the LDS on crop yields (Adejuwon and Odekunle 2006).
In conclusion, on a continental scale, the resemblance of spatial progression patterns thus adds credence to GCM future projections (Tierney et al 2015). However, in regions where the rainfall seasonal cycle is not well captured, caution should be exercised when interpreting climate change projections for impact assessment. The information in figures S24-26 on timing biases for individual models can be used to inform future studies on suitable model selection, reducing the inclusion of models with timing biases.

Acknowledgments, samples, and data
The authors would like to thank two anonymous reviewers for their insightful and constructive comments.
We acknowledge the World Climate Research Programme's Working Group on Coupled Modelling, which is responsible for CMIP, and we thank the climate modeling groups (models listed in table S1) for producing and making available their model outputs; for CMIP, the US Department of Energy's PCMDI provided coordinating support and led development of software infrastructure in partnership with the Global Organization for Earth System Science Portals. Model data were sourced from the CMIP5 data portal (http://cmip-pcmdi.llnl.gov/cmip5/data_portal.html) and the British Atmospheric Data Centre (http:// badc.nerc.ac.uk/).
Caroline M Dunning was supported with funding from a Natural Environment Research Council (NERC) PhD Studentship through the SCENARIO Doctoral Training Partnership grant NE/L002566/1. Richard P Allan's contribution to the research leading to these results has received funding from the National Centre for Earth Observation and the European Union 7th Framework Programme (FP7/2007(FP7/ -2013 under grant agreement 603502 (EU project DACCIWA: Dynamics-aerosol-chemistry-cloud interactions in West Africa). Emily Black's contribution to the research has been supported by the National