A Multi Criteria Decision Analysis Approach for Regional Climate Model Selection and Future Climate Assessment in the Mono River Basin, Benin and Togo

Abstract

Regional climate models (RCMs) are key in the current context of global warming, and they are increasingly used to support decision-making and to identify adaptation measures in response to climate change. However, considering the wide range of available RCMs, it is important to identify the most suitable ones prior to climate impact studies, especially at small scales like catchments. In this study, a multicriteria decision analysis approach, namely the technique for order preferences by similarity to an ideal solution (TOPSIS) was applied to select the best performing RCMs in the Mono River Basin of Benin and Togo (West Africa). The TOPSIS method was used to systematically rank 15 RCMs accessed from the coordinated regional downscaling experiment (CORDEX) database. Six RCMs were finally selected and averaged into an ensemble to assess the future climate in the Mono River Basin until 2070 compared to the period 1966–2015. Two climate change scenarios, RCP 4.5 and RCP 8.5, were considered. The results show that under both climate change scenarios, the annual temperature has an increasing trend during the period 1966–2070, whereas annual rainfall for the next 50 years presents high variability and no statistically significant trend. Furthermore, seasonal cycles of rainfall are expected to change in the different parts of the catchment with delayed onset of rainfall, longer dry seasons, and rainfall intensification. In response to the projected changes, impact studies and risk assessments need to be carried out to evaluate potential implications for human security in the Mono River Basin and to provide adequate adaptation measures.

1. Introduction

Oceans, lands, and the atmosphere have become warmer over the last five decades due to human activities [1]. As a result, changes in the pattern of climate variables and associated consequences such as the rise of mean sea level, ocean acidification, changes in precipitation pattern, and increase in temperature are observed [2]. Climate models, whether global circulation models (GCM) or regional climate models (RCM) are increasingly used to analyse past and future patterns of climate at the global, regional, or local scale. The Intergovernmental Panel on Climate Change (IPCC) bases its assessment reports on scenarios and climate model information disseminated on the Earth System Grid Federation (ESGF) portals. In that regard, a large number of model data are available for the African region within the Coordinated Regional Downscaling Experiment (CORDEX) under the fifth coupled model intercomparison project (CMIP 5).

Due to the coarse resolution and biases embedded in climate models, including both GCMs and RCMs, it is recommended to downscale or bias-correct them before usage in impact studies, especially at small scales like the catchment scale [3]. Among other sources, biases in climate models arise from model parametrization, imperfect conceptualisation, boundary conditions, spatial resolution, and averaging over grids [4,5]. Statistical methods are commonly used for bias correction purposes in regional and local climate studies, as opposed to dynamic downscaling methods which require substantial computational resources [3,6]. Bias correction approaches are built on the assumption that biases remain the same over time, from the past to the future. The correction consists of comparing model historical data against observations to estimate biases that are afterward removed from future datasets [7]. There is a wide range of bias-correction methods that have proven their suitability with respect to different climate variables and depending on the study area. For instance, quantile mapping methods use cumulative distribution functions (CDFs) of observation and historical model data to construct a transfer function used in turn to correct model outputs [8]. The delta change method adds up the difference between observation and model data to adjust biases. It is based on the assumption that changes in climate data are location-specific and occur only over large distances [9,10]. However, due to its simple transfer function, this method does not capture changes in extreme events [11]. The linear scaling method corrects the mean of future data by adjusting the long-term monthly mean of model data to that of the observation [12,13].

In West Africa, data excerpted from the CORDEX database have demonstrated overall good performance in simulating climate in the region. A wide range of local studies based on CORDEX datasets has been carried out in the region over the recent years with satisfactory results. Akinsanola et al., 2015 [14] evaluated the capability of three RCMs, namely REMO, RCA4, and CCLM in simulating West African summer monsoon precipitation and concluded that the first two models simulate rainfall adequately in the region. Likewise, 10 RCMs analysed by Gbobaniyi et. al, 2014 [15] were reported to have acceptable performances in reproducing the spatial distribution of rainfall and temperature over the region. Akinsanola and Ogunjobi, 2017 [16] assessed the performance of RCA4, CRCM5, CCLM, REGCM3, PRECIS, HIRHAM, and REMO against TRMM and CRU rainfall datasets and concluded that they fairly represent the mean annual cycle of rainfall and the interannual variations despite some seasonal and region-specific biases.

Likewise, in the Mono River Basin located in West Africa, future climate assessment studies were carried out during the past years by using various GCMs and RCMs for trend assessment [17,18,19,20], extremes analysis [21,22], and climate change impact studies [23]. However, none of those studies explicitly exposed the selection process of the climate model used. Models were basically selected with reference to other studies where they were reported to be of good performance, or based on data processing constraints, and barely on the basis of a systematic selection. As reported by Browne and Sylla in 2012 [16], the performance of a model within a geographical region like West Africa could vary depending on the location under consideration.

Therefore, the novelty of this study is to carry out a systematic selection of best performing RCMs in the Mono River Basin that will be used afterwards to analyse future climate patterns. In that vein, Lutz et al. [24] have used a three-step process to select best-performing GCMs in the Indus, Ganges, and Brahmaputra river basins. The authors first filtered the GCMs based on their ability to represent changes in mean temperature and rainfall; next, the first selection is refined based on performance vis-à-vis four climate extreme indices; and finally, the second list is trimmed down based on the model’s ability to capture annual cycles. Therefore, this approach basically consists of selecting best-performing models based on predefined criteria, which are evaluated individually. Refaey et al., 2019 [25] furthered that approach by using five statistical metrics with four multicriteria decision analysis (MCDA) techniques to simultaneously evaluate all selected criteria in the Wadi El-Natrun catchment in Egypt. Recently, there has been an increasing interest for MCDA techniques for climate model selection [26,27,28].

Our approach in this paper consists of using a MCDA method to rank RCMs in the Mono River Basin based on statistical and categorical metrics. Furthermore, the RCMs are bias-corrected and ensemble-averaged to assess future climate changes or variation in the next 50 years.

2. Materials and Methods

2.1. Study Area

The Mono River Basin is located within latitudes 6.36° N and 9.39° N, and longitudes 0.62° E and 1.99° E in West Africa (Figure 1a). It stretches over 340 km north to south with an area of 23,592.56 km². The catchment is transboundary and shared by the Republics of Benin and Togo. The Mono River Basin has two main climatic zones defined by a subequatorial climate for latitudes lower than 7° N and a tropical climate in the upper part (latitude above 7° N). Areas of latitude lower than 7° N experience two rainy seasons every year, whereas above 8 °N, the rainfall regime is unimodal with only one peak [29]. Within latitudes 7° N–8° N, the rainfall regime is halfway between a typical unimodal and bimodal cycle, which is described as a “transitional” regime [20]. These three rainfall-based climatic zones are further referred to in this study as south (latitudes < 7° N), centre (7° N < latitude < 8° N), and north (latitudes > 8° N). The lower part of the basin is prone to recurrent flood events which trigger economic losses and deaths in both countries [15]. An average of 1000 mm precipitation per year is recorded in the south and 1200 mm in the northern part [30,31].

2.2. Data

All the RCMs available for the Africa Domain in the CORDEX database and which provide complete time series of rainfall and mean air temperature for RCP 4.5 and RCP 8.5 until 2070 are considered in this study. The list is made of 4 RCMs driven by 8 GCMs. These are in total 15 RCMs that were downloaded from the Earth System Grid Federation (ESGF) node at the German Climate Computing Center (DKRZ) https://esgf-data.dkrz.de/projects/esgf-dkrz/, accessed on 5 March 2020.

Table 1. List of RCMs used and details.

GCM		RCM		GCM-RCM Designation
Name	Developed by	Name	Institute
CNRM-CERFACS-CNRM-CM5	Centre National de Recherches Météorologiques, Centre, France (CNRM)	CCLM4-8-17	Climate Limited-area Modelling Community (CLMcom)	CNRM-CCLM4
ICHEC-EC-EARTH	Irish Centre for High-End Computing (ICHEC)			ICHEC-CCLM4
MOHC-HadGEM2-ES	Met Office Hadley Centre, UK (MOHC)			MOHC-CCLM4
MPI-M-MPI-ESM-LR	Max Planck Institute for Meteorology, Germany (MPI)			MPI-CCLM4
ICHEC-EC-EARTH	ICHEC	RACMO22T	Royal Netherlands Meteorological Institute (KNMI)	ICHEC-RACMO22T
MOHC-HadGEM2-ES	MOHC			MOHC-RACMO22T
CCCma-CanESM2	Canadian Centre for Climate Modelling and Analysis	RCA4	Swedish Meteorological and Hydrological Institute (SMHI)	CCCma-RCA4
CNRM-CERFACS-CNRM-CM5	CNRM			CNRM-RCA4
CSIRO-QCCCE-CSIRO-Mk3-6-0	Commonwealth Scientific and Industrial Research Organization, Australia (CSIRO)			CSIRO-RCA4
IPSL-IPSL-CM5A-MR	Institut Pierre Simon Laplace, France (IPSL)			IPSL-RCA4
MIROC-MIROC5	The University of Tokyo, National Institute for Environmental Studies, and Japan Agency for Marine-Earth Science and Technology, Japan			MIROC-RCA4
MOHC-HadGEM2-ES	MOHC			MOHC-RCA4
MPI-M-MPI-ESM-LR	MPI			MPI-RCA4
ICHEC-EC-EARTH	ICHEC	REMO2009	Helmholtz-Zentrum Geesthacht, Climate Service Center, Max Planck Institute for Meteorology (MPI-CSC)	ICHEC-REMO
MPI-M-MPI-ESM-LR	MPI			MPI-REMO

presents the RCMs used as well as their driving GCM and the designation under which the RCM is subsequently referred to in this study. The RCM data have a spatial resolution of 0.44° × 0.44°, about 50 km × 50 km. Figure 1b presents the distribution of the RCM grids over the catchment. A grid-to-point extraction of the data was performed at location of observation stations in order to facilitate the bias correction process afterwards. In situ data were collected from 38 stations within and around the Mono River Basin (Figure 1a).

For observation data, the mean temperature is given by the average of minimum and maximum temperature collected from met services, whereas for RCMs the mean air surface temperature (named “tas” in the CORDEX database) was downloaded. Observation data cover the period 1966–2015 while 2021–2070 is considered as future period. The overlapping period between observations and model predictions for the past is 1966–2005.

2.3. Ranking and Selection of RCMs

The selection of RCMs to be used for future climate assessment was based on the TOPSIS method, a technique for order preferences by similarity to an ideal solution (Hwang and Yoon, 1981). TOPSIS is a multicriteria decision-making approach for sorting alternatives based on a compromise solution. The best alternative is identified as the closest to the positive ideal solution and the farthest from the negative ideal solution. The TOPSIS method is widely used for ranking alternatives and for decision making in water resources management, early warning systems, participatory flood risk management, social learning, and consensus achievement among stakeholders [32,33,34,35]. It has been increasingly used in the last decade in the field of climatology to select among different datasets [24,25,26,27,28,36]. The backbone of the TOPSIS method is the existence of many alternatives that are ordered based on criteria.

Considering a set of alternatives $A_k$, $k=1,\dots ,n$, a set of criteria $C_j, j=1,\dots ,m$, $x_{kj}$ the performance ratings of alternative $k$ to criteria $j$, and $w_j$ the weight attributed to each criteria, the TOPSIS approach consists of the following steps.

• Normalization of performance ratings.

For criteria to maximise, also called benefit criteria (the larger, the better), the normalized rating $r_{kj}$ is given by

$$r_{kj}\left(x\right)=\frac{x_{kj}-x_j^{-}}{x_j^{*}-x_j^{-}}, k=1,\dots , n; j = 1,\dots , m,$$

(1)

where $x_j^{*}$ is the aspired/desired level of criteria $j$ and $x_j^{-}$ its the worst level.

For criteria to minimise or cost criteria (the smaller, the better), the normalized rating is given by

$$r_{kj}\left(x\right)=\frac{x_j^{-}-x_{kj}}{x_j^{-}-x_j^{*}}.$$

(2)

• Calculation of weighted normalized ratings, $v_{kj}\left(x\right)$:

$$v_{kj}\left(x\right)=w_j{r}_{kj}\left(x\right), k=1,\dots , n; j = 1,\dots , m.$$

(3)

• Derivation of positive ideal solution (PIS) and negative ideal solution (NIS).

Because there is no good or bad alternative, $PIS$ and $NIS$ represent respectively the most preferable and the less desired set of criteria one wish to achieve. $PIS$ and $NIS$ are given by

$$PIS=\left\{\left(\underset{k}{\max }{v}_{kj}\left(x\right)\mid j \in J_1\right),\left(\underset{k}{\min }{v}_{kj}\left(x\right)\mid j \in J_2\right),k=1,\dots ,n\right\}$$

(4)

$$NIS=\left\{\left(\underset{k}{\min }{v}_{kj}\left(x\right)\mid j \in J_1\right),\left(\underset{k}{\max }{v}_{kj}\left(x\right)\mid j \in J_2\right),k=1,\dots ,n\right\},$$

(5)

where $J_1$ and $J_2$ are the benefit and cost elements respectively.

• Estimation of separation from the PIS and the NIS.

The separation from the $PIS$, $D_k^{+}$, and from the $NIS$, $D_k^{-}$ can be estimated as Euclidean distance with Equations (6) and (7):

$$D_k^{+}=\sqrt{{{\displaystyle \sum }}_{j=1}^m{\left[v_{kj}\left(x\right)-v_j^{+}\left(x\right)\right]}^2},k=1,\dots ,n$$

(6)

$$D_k^{-}=\sqrt{{\sum }_{j=1}^m{\left[v_{kj}\left(x\right)-v_j^{-}\left(x\right)\right]}^2},k=1,\dots ,n.$$

(7)

• Derivation of similarities to the PIS.

The similarities to the $PIS$ are computed as

$$C_k^{+}=\frac{D_k^{-}}{D_k^{+}+D_k^{-}},k=1,\dots ,n.$$

(8)

• Ordering of alternatives according to the similarities to PIS in a decreasing order.

Finally, the alternatives can be ranked from most preferred to less preferred by ordering $C_k^{+}$ in decreasing order.

In this study, daily rainfall and temperature from RCMs are the alternatives and the criteria consist of a set of statistical metrics. These metrics are the Nash–Sutcliffe efficiency (NSE), the coefficient of determination R², and two categorical metrics: the probability of detection (POD) and the false alarm ratio (FAR). POD and FAR are specifically applied to rainfall. POD represents the likelihood for RCMs to detect a rainfall event, and the false alarm ratio (FAR) describes the fraction of predicted rainfall event that did not actually happen [37]. We have

$$NSE=1-\frac{{\sum }_{i=1}^n{\left(O_i-S_i\right)}^2}{{\sum }_{i=1}^n{\left(O_i-\overline{O}\right)}^2}$$

(9)

$$R^2={\left(\frac{{\sum }_{i=1}^n\left(O_i-\overline{O}\right)\left(S_i-\overline{S}\right)}{\sqrt{{\sum }_{i=1}^n{\left(O_i-\overline{O}\right)}^2}\sqrt{{\sum }_{i=1}^n{\left(S_i-\overline{S}\right)}^2}}\right)}^2,$$

(10)

where O_i and S_i are respectively observed and model values at time i, $\overline{O}$ and $\overline{S}$ are respectively the mean observed and model values, and n the sample size,

$$POD=\frac{hits}{hits+misses}$$

(11)

$$FAR=\frac{false alarms}{hits+false alarms},$$

(12)

where hits is the number of rainfall events that are effectively predicted by the RCM, misses is the the number of observed rainfall events that were not predicted by the RCM, and false alarms is the events predicted by RCMs but did not actually occur.

More details on these metrics are provided by [30,38]. All four criteria were weighted equally with weight 0.25 for rainfall, whereas NSE and R² were weighted each 0.5 for temperature.

The ranking approach with TOPSIS was applied to both precipitation and temperature datasets at each station location to determine the performance of each RCM in different parts of the catchment. A heatmap was used to visualise the number of locations at which a given RCM ranked $k^{th}, k=1,\dots n$. As this study uses 15 RCMs, the heatmap has a 15 by 15 dimension.

The overall performance of each RCM respective to the whole catchment (not at individual station locations) was determined based on their occurrence frequency at different locations. Following the method of Homsi et al., 2020 [26], a score is computed by using the ranks of RCMs at individual locations and their frequency of occurrence. If an RCM had a rank of $1,2,3,\dots ,n_z$ respectively at $l_1, l_2, l_3,\dots l_z$ locations, the score of that RCM is given by $\sum _z\left(\frac{l_z}{n_z}\right)$. Therefore, the higher the occurrence frequency of a RCM, as well as its rank at individual location, the more weight it is assigned, and therefore the higher its overall rank compared to the other RCMs. For instance, if an RCM got rank 1 at $p$ locations, rank 2 at $q$ locations, …, rank 15 at $s$ locations, then its overall score in the catchment is given by: $\frac{p}{1}+\frac{q}{2}+\dots +\frac{s}{15}$.

Finally, the most suitable models for making an ensemble are defined as those falling in the upper 50th percentile of all RCMs, for both precipitation and temperature [26,39].

2.4. Bias Correction

To bias-correct the RCMs, the quantile mapping approach, also called the quantile–quantile method or distribution mapping was applied for precipitation and temperature datasets. The quantile mapping was used considering its good results in different climatic zones all over the world [12,40,41,42,43,44,45,46], and in other West African catchments similar to the Mono River Basin [8,47,48,49]. Furthermore, previous studies in the Mono River Basin have reported good performances with the quantile method [17,18,20]. Quantile mapping methods use cumulative distribution functions (CDFs) of observation and historical model data to construct a transfer function used in turn to correct model outputs [8]. Generally, in the application of the quantile mapping method, the Gamma distribution (Equation (13)) is used for precipitation and the Gaussian for temperature (Equation (14)) [45]. We have

$$f_{\gamma }\left(x_{\alpha ,\beta }\right)=x^{\alpha -1}·\frac{1}{{\beta }^{\alpha }·\gamma \left(\alpha \right)}·e^{\frac{-x}{\beta }};x\ge 0;\alpha ,\beta >0,$$

(13)

where $\alpha$ is the shape parameter and $\beta$ the scale parameter, and

$$f_N\left(x_{\mu ,{\sigma }^2}\right)=x^{\alpha -1}·\frac{1}{\sigma .\sqrt{2\pi }}·e^{\frac{-{\left(x-\mu \right)}^2}{2{\sigma }^2}};x\in R,$$

(14)

with µ and σ the location and scale parameters respectively.

2.5. Future Climate Trend Assessment

To visualise the spatial distribution over past and future periods, the kriging interpolation method [20] was used. The past period is defined as 1966–2015 and the future as 2021–2070. The percentage of change (Equation (15)) was computed to estimate future changes with respect to the past. We have

$$P_V=\frac{V_{proj}-V_{obs}}{V_{obs}}\times 100,$$

(15)

where, $P_V$ is the percentage of change, $V_{proj}$ is the average value of variable $V$ for the future period, and $V_{obs}$ the average value during observation period.

In addition, the Mann–Kendall test [50] at 95% confidence interval was used to analyse trends in annual rainfall accumulation and mean annual temperature. A positive Z value from the Mann–Kendall test corresponds to an increasing trend (a negative value to a decreasing trend) and a value lower than 1.96 indicates that the trend is statistically not significant. Sens’ slope was also computed to estimate magnitudes of increase or decrease.

3. Results

3.1. Ranking and Selection of RCMs

3.1.1. TOPSIS Results: Best RCM per Location

The TOPSIS analysis provided a ranking of RCMs at the different locations considered across the catchment. Based on the TOPSIS scores presented in Appendix A (

Table A1. Topsis ranking scores for rainfall.

Station	CNRM- CCLM4	ICHEC- CCLM4	MOHC- CCLM4	MPI- CCLM4	ICHEC-RACMO22T	MOHC- RACMO22T	CCCma- RCA4	CNRM- RCA4	CSIRO- RCA4	IPSL- RCA4	MIROC- RCA4	MOHC- RCA4	MPI-RCA4	ICHEC-REMO	MPI-REMO
Abomey	0.03	0.84	0.83	0.87	0.87	0.13	0.10	0.76	0.25	0.16	0.88	0.94	0.94	0.84	0.27
Adeta	0.12	0.87	0.84	0.84	0.83	0.11	0.10	0.11	0.22	0.08	0.25	0.28	0.89	0.87	0.22
Afagnan	0.01	0.10	0.11	0.11	0.11	0.08	0.06	0.06	0.11	0.07	0.14	0.16	0.17	0.11	0.90
Agouna	0.03	0.51	0.52	0.93	0.92	0.49	0.06	0.11	0.16	0.09	0.52	0.21	0.53	0.88	0.18
Akaba	0.36	0.87	0.61	0.62	0.62	0.08	0.32	0.13	0.20	0.11	0.39	0.21	0.64	0.64	0.34
Aklakou	0.10	0.16	0.18	0.18	0.17	0.16	0.16	0.16	0.21	0.18	0.24	0.26	0.27	0.18	0.81
Amou-Oblo	0.18	0.49	0.74	0.76	0.53	0.49	0.12	0.18	0.22	0.16	0.26	0.55	0.48	0.55	0.47
Aneho-Glidji	0.11	0.20	0.19	0.23	0.77	0.23	0.21	0.18	0.27	0.25	0.29	0.34	0.35	0.66	0.96
Anie-Mono	0.26	0.91	0.60	0.93	0.30	0.45	0.46	0.53	0.33	0.19	0.59	0.59	0.94	0.35	0.52
Aplahoue	0.02	0.80	0.83	0.85	0.89	0.12	0.10	0.14	0.22	0.15	0.87	0.96	0.94	0.82	0.25
Atakpame	0.10	0.91	0.53	0.53	0.90	0.49	0.10	0.13	0.19	0.49	0.53	0.53	0.54	0.90	0.23
Athieme	0.03	0.14	0.14	0.87	0.91	0.13	0.13	0.79	0.18	0.14	0.20	0.25	0.25	0.83	0.18
Bante	0.47	0.85	0.92	0.91	0.53	0.49	0.46	0.17	0.23	0.09	0.54	0.53	0.93	0.53	0.53
Bassila	0.34	0.84	0.91	0.90	0.90	0.33	0.33	0.12	0.21	0.09	0.39	0.41	0.94	0.86	0.40
Blitta	0.37	0.89	0.81	0.83	0.67	0.61	0.33	0.34	0.19	0.07	0.37	0.41	0.87	0.94	0.34
Bohicon	0.02	0.52	0.51	0.52	0.84	0.10	0.09	0.49	0.18	0.11	0.52	0.54	0.54	0.84	0.21
Bopa	0.02	0.16	0.17	0.18	0.89	0.12	0.11	0.10	0.19	0.14	0.91	0.25	0.25	0.83	0.20
Dogbo	0.03	0.11	0.10	0.11	0.11	0.08	0.07	0.06	0.12	0.07	0.95	0.16	0.16	0.12	0.12
Grand-Popo	0.05	0.09	0.10	0.12	0.17	0.16	0.14	0.11	0.17	0.15	0.18	0.22	0.22	0.10	0.91
Kara	0.38	0.89	0.89	0.90	0.41	0.29	0.30	0.38	0.32	0.30	0.46	0.40	0.92	0.46	0.65
Kougnohou	0.19	0.82	0.63	0.65	0.53	0.31	0.30	0.31	0.24	0.18	0.33	0.38	0.67	0.77	0.27
Kpewa-Aledjo	0.31	0.90	0.75	0.91	0.66	0.29	0.28	0.30	0.23	0.29	0.36	0.35	0.92	0.78	0.54
Lokossa	0.02	0.17	0.89	0.86	0.82	0.79	0.17	0.12	0.23	0.17	0.22	0.27	0.30	0.80	0.25
Lonkly	0.03	0.84	0.87	0.87	0.86	0.11	0.08	0.14	0.19	0.10	0.24	0.26	0.95	0.21	0.21
Malfacassa	0.35	0.66	0.68	0.91	0.68	0.34	0.33	0.34	0.20	0.34	0.40	0.22	0.94	0.67	0.38
Nangbeto	0.05	0.86	0.93	0.90	0.79	0.15	0.11	0.06	0.26	0.07	0.80	0.89	0.31	0.84	0.29
Niaouli	0.03	0.12	0.12	0.13	0.94	0.10	0.09	0.07	0.15	0.11	0.16	0.20	0.20	0.87	0.15
Notse	0.08	0.88	0.18	0.92	0.15	0.09	0.07	0.10	0.16	0.07	0.19	0.22	0.98	0.16	0.18
Penesoulou	0.27	0.90	0.76	0.76	0.53	0.27	0.26	0.28	0.17	0.25	0.31	0.30	0.53	0.52	0.32
Savalou	0.07	0.87	0.52	0.92	0.52	0.49	0.09	0.09	0.17	0.08	0.50	0.52	0.53	0.88	0.18
Sokode	0.28	0.91	0.53	0.93	0.92	0.28	0.28	0.28	0.19	0.10	0.32	0.33	0.75	0.75	0.33
Sotouboua	0.13	0.93	0.59	0.59	0.43	0.22	0.22	0.22	0.16	0.06	0.25	0.28	0.78	0.62	0.23
Tabligbo	0.01	0.17	0.18	0.89	0.83	0.12	0.12	0.10	0.19	0.14	0.19	0.24	0.25	0.81	0.23
Tchamba	0.29	0.91	0.74	0.90	0.32	0.27	0.26	0.26	0.30	0.07	0.33	0.32	0.75	0.53	0.50
Tchetti	0.04	0.84	0.88	0.90	0.85	0.79	0.08	0.14	0.19	0.10	0.86	0.24	0.93	0.83	0.22
Toffo	0.04	0.17	0.13	0.16	0.90	0.08	0.06	0.10	0.15	0.11	0.19	0.22	0.21	0.88	0.89
Wahala	0.11	0.87	0.24	0.91	0.89	0.11	0.09	0.13	0.21	0.11	0.23	0.27	0.97	0.88	0.24
Yegue	0.34	0.66	0.68	0.91	0.40	0.63	0.33	0.35	0.19	0.10	0.40	0.40	0.93	0.67	0.40

) and Appendix B (

Table A2. Topsis ranking scores for temperature.

Station	CNRM- CCLM4	ICHEC- CCLM4	MOHC- CCLM4	MPI- CCLM4	ICHEC- RACMO22T	MOHC- RACMO22T	CCCma- RCA4	CNRM- RCA4	CSIRO- RCA4	IPSL- RCA4	MIROC- RCA4	MOHC- RCA4	MPI- RCA4	ICHEC- REMO	MPI- REMO
Abomey	0.34	0.53	0.80	0.83	0.75	0.37	0.43	0.39	0.53	0.46	0.08	0.21	0.76	0.62	0.76
Adeta	0.23	0.75	0.67	0.63	0.56	0.31	0.46	0.26	0.43	0.31	0.00	0.18	0.89	0.46	0.64
Afagnan	0.38	0.49	0.75	0.84	0.75	0.35	0.41	0.42	0.56	0.48	0.00	0.22	0.67	0.67	0.82
Agouna	0.29	0.24	0.89	0.74	0.76	0.47	0.38	0.38	0.62	0.47	0.20	0.33	0.47	0.57	1.00
Akaba	0.52	0.22	0.58	0.31	0.30	0.21	0.24	0.66	0.83	0.72	0.48	0.56	0.27	0.18	0.74
Aklakou	0.44	0.31	0.67	0.72	0.50	0.50	0.53	0.51	0.65	0.56	0.03	0.29	0.42	0.53	0.94
AmouOblo	0.46	0.34	0.68	0.41	0.39	0.09	0.18	0.59	0.79	0.67	0.37	0.48	0.41	0.23	0.86
AnehoGlidji	0.44	0.31	0.67	0.72	0.50	0.50	0.53	0.51	0.65	0.56	0.03	0.29	0.42	0.53	0.94
AnieMono	0.41	0.14	0.69	0.49	0.50	0.38	0.30	0.55	0.70	0.60	0.38	0.45	0.27	0.38	0.86
Aplahoue	0.34	0.58	0.77	0.83	0.75	0.37	0.44	0.40	0.54	0.46	0.08	0.21	0.82	0.63	0.76
Atakpame	0.31	0.21	0.75	0.64	0.63	0.39	0.31	0.43	0.62	0.50	0.21	0.33	0.41	0.49	0.89
Athieme	0.38	0.49	0.75	0.84	0.75	0.35	0.41	0.42	0.56	0.48	0.00	0.22	0.67	0.67	0.82
Bante	0.40	0.29	0.56	0.65	0.53	0.41	0.40	0.46	0.67	0.52	0.12	0.42	0.45	0.51	0.51
Bassila	0.45	0.28	0.55	0.63	0.48	0.32	0.33	0.51	0.68	0.56	0.23	0.53	0.43	0.49	0.49
Blitta	0.52	0.34	0.45	0.43	0.34	0.31	0.37	0.58	0.76	0.67	0.25	0.56	0.39	0.28	0.38
Bohicon	0.34	0.53	0.80	0.83	0.75	0.37	0.43	0.39	0.53	0.46	0.08	0.21	0.76	0.62	0.76
Bopa	0.38	0.49	0.75	0.84	0.75	0.35	0.41	0.42	0.56	0.48	0.00	0.22	0.67	0.67	0.82
DogboTota	0.38	0.49	0.75	0.84	0.75	0.35	0.41	0.42	0.56	0.48	0.00	0.22	0.67	0.67	0.82
GrandPopo	0.59	0.53	0.65	0.62	0.43	0.12	0.17	0.62	0.68	0.68	0.37	0.62	0.56	0.47	0.46
Kara	0.59	0.53	0.65	0.62	0.43	0.12	0.17	0.62	0.68	0.68	0.37	0.62	0.56	0.47	0.46
Kougnohou	0.46	0.34	0.68	0.41	0.39	0.09	0.18	0.59	0.79	0.67	0.37	0.48	0.41	0.23	0.86
KpewaAledjo	0.50	0.33	0.54	0.58	0.39	0.11	0.16	0.60	0.64	0.64	0.36	0.68	0.45	0.39	0.42
Lokossa	0.38	0.49	0.75	0.84	0.75	0.35	0.41	0.42	0.56	0.48	0.00	0.22	0.67	0.67	0.82
Lonkly	0.34	0.58	0.77	0.83	0.75	0.37	0.44	0.40	0.54	0.46	0.08	0.21	0.82	0.63	0.76
Malfacassa	0.40	0.23	0.54	0.60	0.44	0.25	0.25	0.52	0.57	0.55	0.24	0.60	0.43	0.42	0.44
Nangbeto	0.31	0.21	0.75	0.64	0.63	0.39	0.31	0.43	0.62	0.50	0.21	0.33	0.41	0.49	0.89
Niaouli	0.38	0.49	0.78	0.85	0.75	0.32	0.39	0.42	0.57	0.48	0.00	0.22	0.69	0.67	0.82
Notse	0.24	0.66	0.80	0.89	0.80	0.44	0.54	0.28	0.45	0.33	0.00	0.18	0.94	0.68	0.80
Penesoulou	0.50	0.29	0.57	0.61	0.45	0.21	0.24	0.57	0.66	0.62	0.32	0.59	0.44	0.47	0.48
Savalou	0.31	0.17	0.82	0.60	0.63	0.39	0.33	0.45	0.65	0.51	0.26	0.34	0.34	0.49	0.97
Sokode	0.49	0.27	0.50	0.55	0.39	0.27	0.29	0.55	0.66	0.61	0.25	0.57	0.40	0.39	0.42
Sotouboua	0.52	0.34	0.45	0.43	0.34	0.31	0.37	0.58	0.76	0.67	0.25	0.56	0.39	0.28	0.38
Tabligbo	0.38	0.51	0.75	0.85	0.75	0.35	0.43	0.42	0.57	0.49	0.00	0.22	0.73	0.68	0.77
Tchamba	0.50	0.33	0.54	0.58	0.39	0.11	0.16	0.60	0.64	0.64	0.36	0.68	0.45	0.39	0.42
Tchetti	0.29	0.24	0.89	0.74	0.76	0.47	0.38	0.38	0.62	0.47	0.20	0.33	0.47	0.57	1.00
Toffo	0.34	0.53	0.80	0.83	0.75	0.37	0.43	0.39	0.53	0.46	0.08	0.21	0.76	0.62	0.76
Wahala	0.24	0.33	0.89	0.88	0.86	0.47	0.33	0.30	0.60	0.38	0.18	0.39	0.57	0.70	1.00
Yegue	0.44	0.25	0.50	0.68	0.53	0.41	0.38	0.51	0.68	0.59	0.16	0.48	0.44	0.48	0.43

), the RCMs which ranked first were derived and mapped as shown on Figure 2.

The spatial distribution of rank 1 RCMs for temperature shows a pattern whereby MPI-CCLM4 predominantly performed best in the south, MPI-REMO in the centre, and CSIRO-RCA4 in the north. As for rainfall, there is no spatial pattern. However, MPI-RCA4 is the most present all over the catchment from north to south. Similar results with spatial pattern were reported by Homsi et al., 2020 [26]. The authors noticed a spatial pattern dominated by three GCMs for rainfall, whereas temperature displayed a different distribution of five new GCMs mixed with only one from the bests of rainfall. Spatial patterns in best performing models were also found by Shiru et al., 2019 [27] in Nigeria.

An RCM holding the first position in a certain location does not make it a suitable model at catchment scale because it may perform poorly elsewhere in the study area [51]. Information on the first ranking RCM at an individual location can actually be useful for local studies in different parts of the catchment. Moreover, it points out the fact that RCMs perform differently depending on the climate variable and from one location to another. For example, MPI-CCLM4 ranked first in most areas of the southern part for temperature but came first only at one location for rainfall. Nonetheless, the model performed relatively well, occupying rank 2 at 13 locations (Figure 3b).

Figure 3a,b are heatmaps displaying the number of times (at how many locations) a given RCM occupied a certain rank. The darker the color of a cell, the higher the number of stations where an RCM occupies the rank corresponding to the cell. For instance, for temperature (Figure 3a), MIROC-RCA4 occupied rank 15 at 25 stations of 38. Likewise, Figure 3b shows that CNRM-CCLM4 occupied rank 15 at 20 locations over the 38 stations considered, and similarly, CCCma-RCA4 and IPSL-RCA4 occupied only rank 8 onward, making these three RCMs the ones with lowest performances for rainfall.

3.1.2. RCMs Selection

Table 2. Overall ranking of RCMs for rainfall and temperature.

Rank	Rainfall		Temperature
	Model	Score	Model	Score
1	MPI-RCA4	20.23	MPI-CCLM4	19.99
2	MPI-CCLM4	15.59	MPI-REMO	18.21
3	ICHEC-CCLM4	12.40	CSIRO-RCA4	16.03
4	ICHEC-RACMO22T	11.39	MOHC-CCLM4	12.6
5	ICHEC-REMO	11	IPSL-RCA4	8.68
6	MOHC-CCLM4	9.67	MPI-RCA4	8.06
7	MOHC-RCA4	9.02	MOHC-RCA4	7.18
8	MPI-REMO	8.73	ICHEC-RACMO22T	6.95
9	MIROC-RCA4	7.89	CNRM-RCA4	5.79
10	CSIRO-RCA4	3.96	ICHEC-REMO	5.09
11	MOHC-RACMO22T	3.77	ICHEC-CCLM4	4.05
12	CNRM-RCA4	3.41	CNRM-CCLM4	3.94
13	IPSL-RCA4	3.08	CCCma-RCA4	3.54
14	CCCma-RCA4	2.97	MOHC-RACMO22T	3.20
15	CNRM-CCLM4	2.94	MIROC-RCA4	2.76

presents the overall ranking of RCMs for rainfall and temperature in the Mono catchment. RCMs perform differently vis-à-vis the two variables, e.g., MPI-RCA4 ranked first for rainfall and has rank 6 for temperature. Finally, to make the ensemble of models to be used for future climate assessment, only RCMs whose ranks simultaneously fall within the range 1–8 for both temperature and rainfall are selected. These are six RCMs, highlighted in bold in Table 2: MPI-RCA4, MPI-CCLM4, ICHEC-RACMO22T, MOHC-CCLM4, MOHC-RCA4, and MPI-REMO. All three RCMs driven by the MPI global model are part of the final list. In fact, good performances of the MPI GCM have been reported in other catchments in Benin and Togo [47,48,52] and in the Mono River Basin [20,21]. Because boundary conditions of RCMs are provided by their driving GCMs [53], the high ranking of MPI-driven RCMs indicates that those RCMs may better represent local climate in the Mono catchment.

However, an RCM driven by a certain GCM that performs well does not guarantee good results of the GCM–RCM model because biases in models arise from both RCM and the driving GCM [21]. Furthermore, all four types of RCM in the initial list of models considered in this study (RCA4, CCLM4, RACMO22T, and REMO) are actually represented in the final list of shortlisted models. Because models’ performances are region- and variable-specific, composing the average ensemble with models that performed well for the two climate variables under consideration and over the entire catchment, increases the opportunity to capture actual climate patterns in the study area. In fact, model ensembles improve on individual models performances and even outperform them [15,47].

3.2. Assessment of Future Climate

The evaluation of future climate state is based on the mean ensemble of the six best performing RCMs identified above. The mean of those RCMs was computed for temperature and for rainfall to make the ensemble for each variable. The evaluation was conducted through visualisation of variables’ spatial distribution, quantification of relative changes, and annual trend assessment.

3.2.1. Temperature

Figure 4 presents the spatial distribution of average annual temperature for the observation period 1966–2015, and for the future period 2021–2070 under climate scenarios RCP 4.5 and RCP 8.5. An increase of temperature is expected all over the catchment, but more specifically in the northern part and in the downstream area (south) (Figure 4b,c). However, RCP 8.5 projects warmer conditions than RCP 4.5.

The increase in temperature is also depicted by the annual trend 1966–2070 as shown in Figure 5.

From an average annual temperature of 26.9 °C during the period 1966–2015, it is expected to reach 27.8 °C under RCP 4.5 and 28.4 °C for RCP 8.5. Thus, the increase of average annual temperature over the Mono catchment during the period 2021–2070 is estimated to be 1 °C and 1.5 °C under the medium and high pathway scenarios used in this study. This is supported by the Mann–Kendall test and Sens’ slope results presented in

Table 3. Results of Mann–Kendall and Sens’ slope tests on annual temperature.

Scenario	Z Statistics	p-Value	Sens’ Slope
RCP 4.5	6.67	0.00	0.04
RCP 8.5	7.81	0.00	0.06

.

The results are in line with the graphical observation and confirm that the trend in future temperature is statistically significant with 95% confidence. Moreover, the Sens’ slope values indicate that annual temperature will increase by 0.04 °C every year based on the scenario RCP 4.5 and 0.06 °C according to projections by RCP 8.5. Overall, there is an agreement of both scenarios on the trend of temperature in the Mono catchment for the future period 2021–2070. Regardless of the models used, previous studies in the Mono catchment have also reported an increase of temperature at horizon 2050 up to 2100 [18,20,22]. Similar trends are also found across Africa [54,55,56] and are in line with global patterns predicted by the IPCC [57].

3.2.2. Rainfall

Figure 6 shows the changes in annual rainfall distribution over the Mono river catchment from 1966–2015 to 2021–2070. Figure 6b,c displays similar changes for both RCP 4.5 and RCP 8.5. The western and northwestern parts of the catchment are expected to receive less precipitation in the future, whereas the east central part and the downstream area close to the outlet might experience an increase of annual precipitation. The central part is a mountainous region; thus, the increase may be due to local orographic terrain [21].

As for the downstream area, it is located near the coast where evaporation and cloud formation above the Atlantic ocean may induce the increase of precipitations [53,58]. However, as reported by Lischeid et al. (2021) [59], trend analysis in climate change studies may be affected by artefacts in local data. The authors analysed trends of water level and groundwater head data in Northeast Germany and found that the apparent inconsistent trends observed could be attributed to low-pass filtering of the groundwater recharge signal. As reported in the fourth assessment report of the IPCC, artefacts in the models’ data are addressed for instance with low-pass filters but may still persist at local scale [60].

Overall, average annual rainfall over the Mono catchment depicts high interannual variabilities as shown in Figure 7.

Based on the climate change scenarios, the average total rainfall per year for the future period will be 1091 mm for RCP 4.5 and 1053 mm for RCP 8.5. With both scenarios, the mean annual rainfall is expected to be less than the 1195 mm recorded during 1966–2015. The minimum annual rainfall for the next five decades is expected to be lower (697 mm for RCP 4.5 and 757 mm for RCP 8.5) than the minimum recorded during the past 50 years (854 mm). The high interannual variability observed graphically was confirmed by the Mann–Kendall test (

Table 4. Results of Mann–Kendall and Sens’s slope tests on annual rainfall.

Scenario	Z Statistics	p-Value	Sens’ Slope
RCP 4.5	0.03	0.97	0.1
RCP 8.5	−1.54	0.12	−2.94

). The results show no statistically significant trend (p-values > 0.05), whereas the Sens’ slope estimator indicates an average increase (decrease) of 0.1 mm (2.94 mm) per year with the scenario RCP 4.5 (RCP 8.5) during the period 2021–2070.

Such high variability and insignificant trend in annual precipitation have been reported in previous studies [18,20] in the Mono catchment. Moreover, the mean annual cycle indicates changes in seasons and in monthly precipitations (Figure 8). Overall, the start of the rainy seasons is likely to be delayed all over the catchment. For instance, the first rainfall events that usually occur in March in the south, are likely to shift in April based on the projections by RCP 4.5 and RCP 8.5. Moreover, the amounts of rainfall recorded in April and May in the past are expected to decrease in the future, according to the climate change scenarios. The bimodal rainfall cycle in the south will become “transitional” under the scenario RCP 4.5 and unimodal under RCP 8.5, whereas the central part will shift to a unimodal regime in the future. Such changes would lead to modifications in the agricultural calendar of the concerned agro-ecological zones and can be detrimental for crop production and for food security.

There is an increase of peaks during rainy seasons and a decrease of precipitation during the dry season. Therefore, rainy seasons may become wetter and dry seasons dryer compared to the past. Receiving higher amounts of precipitation during a shorter period of time will translate into rainfall intensification which may increase flood risk and the probability of extreme events in the area [46]. The percentage of future changes in the annual cycle with respect to the past is presented in Figure 9.

Figure 9 indicates that highest increases are expected in the south during July and August and more specifically in August where a greater than 100% increase was found. These projections need to be given attention because the period of July and August in the southern part is usually characterised by a rainfall cessation time during which farmers harvest and prepare the land for the second sowing season of the year [61]. Furthermore, the highest decreases of precipitation, 93–100%, are expected to occur from December to March all over the catchment. In the north, almost all months, except September and October, showed a reduction of rainfall amounts.

Historically, the Mono catchment experiences recurrent flood events [62]. Nonetheless, considering the potential decrease of rainfall (even statistically not significant), the increase of temperature discussed above, and the changes in future land use/cover mainly characterized by a savannahfication of forests and agricultural lands [31], drought-related studies should be undertaken alongside flood assessments in the Mono river catchment.

4. Conclusions

This study assessed future climate conditions in the Mono River Basin by using the TOPSIS multicriteria decision method to identify best-performing regional climate models. RCMs were ranked based on four statistical and categorical metrics (NSE, R², POD, FAR) applied at 38 measurement stations. Finally, six RCMs were selected out of the initial list of 15 to make an ensemble that is further used to evaluate potential changes in future climate by 2070 and under climate change scenarios RCP 4.5 and RCP 8.5. Both scenarios suggest a 1 to 1.5 °C increase of annual temperature in the catchment, especially in the northern and southern parts. On the other hand, a statistically insignificant decreasing trend was found in annual precipitation. The seasonal cycle of rainfall during the period 2021–2070 will be characterized by shorter rainy seasons and an increase of precipitation. This intensification of rainfall may exacerbate existing flood risks in the Mono River Basin. However, the compound effect of temperature rise, dry season prolongation and land use/cover changes may introduce drought as a major hazard in the study area in addition to floods. Moreover, the concordance between the results of the two climate change scenarios used indicates a relatively high possibility that the projections actually occur. Nevertheless, the possibilities of potential artefacts in models can be investigated and addressed by future studies in the Mono catchment. The findings of this study should be furthered by assessing the impact of the projected changes on flood, drought, agriculture, and health to support decision making and the identification of appropriate adaptation measures.