Burkina Faso (BF) is a landlocked country in West Africa. It is located between 9˚20' and 15˚05' North latitude, 5˚30' West longitude, and 2˚20' East longitude ( Figure 1 ). It is bordered by six countries: Mali to the north, Niger to the east, Benin to the southeast, Togo and Ghana to the south, and Côte d’Ivoire to the southwest. Located in the Sahel region, it experiences a distinct wet season and dry season driven by the West African Monsoon [29] .

The timing and intensity of the monsoon strongly affect rainfall patterns during the wet season. The aforementioned occurs from May to October and lasts three months in the North and up to six months in the South. The annual cumulative rainfall varies between 300 mm in the North and less than 1200 mm [30] with three climatic zones: Sahelian zone (rainfall less than 300 mm), Soudano-sahelian zone (rainfall between 300 mm and 900 mm), and Soudanian zone (rainfall greater than 900 mm). The irregular distribution of rainfall has consequences for rainfed agriculture. In fact, the observed changes in planting dates, the occurrence of long dry spells, and the limited water availability for crop growth are the climate-related limiting factors for food production [31] .

Crop models are developed to address various challenges and objectives in the field of crop production and management. Efforts have led to the development of more advanced models, some of which are more focused on the plot-plant scale, while others, like AquaCrop, are more focused on the canopy-level scale and serve as management tools to assist in decision making.

The AquaCrop model has been developed by the Food and Agriculture Organization (FAO) of the United Nations [32] . It is specifically designed to simulate and optimize crop production, with a particular emphasis on water management in agriculture, particularly for water-sensitive crops [32] . The model considers various factors, such as planting date, climate, available soil water, field management, and crop-specific parameters, in order to simulate crop growth and development. Climate data used in the model typically includes daily rainfall, minimum (Tn) and maximum (Tx) air temperature, reference crop evapotranspiration (ETo), and mean annual carbon dioxide concentration (CO₂). By making adjustments to parameters such as planting date, agronomic practices, and crop-specific location parameters, users can determine the most efficient strategies for enhancing crop yields and water use efficiency [33] . The following AquaCrop chart illustrates the key elements of the soil-plant-atmosphere continuum and the parameters that influence phenology, canopy cover, transpiration, biomass production, and yield ( Figure 2 ). In this study, the open-source AquaCrop Fortran code from the FAO website (available at https://www.fao.org/aquacrop/software/aquacrop-gisen#c518675 ) has been compiled on a Unix-like operating system and then used for simulation at the grid cell resolution of 0.44˚ × 0.44˚ across BF.

Daily climate data for the period 1983-2020 are used in this study, including precipitation, minimum and maximum temperature, dew point temperature at 2 m, and surface net solar radiation. Except for precipitation data, climate data were retrieved from ERA5, the fifth generation of global climate reanalysis data from the European Centre for Medium-Range Weather Forecasts [35] [36] . The data are accessible on Copernicus Climate Data Store ( https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels?tab=overview ). ERA5 has been widely validated and successfully applied in agricultural climate impact studies [37] [38] .

Precipitation data is obtained by merging measurements from rain gauges with satellite precipitation data. In Burkina Faso, the rain gauge network consists of nearly 134 observation stations, which include synoptic, climatological, agrometeorological, and rain gauges that have long-term series of data. To ensure data quality, stations with large data gaps (greater than 5% on a daily basis) for the considered period are removed, resulting in 97 rain gauges for the merging process ( Figure 3 ). Satellite data are from the version 3.1 of the Tropical Applications of Meteorology using SATellite data and ground-based observations (TAMSAT v3.1) [39] . Previous research comparing seven satellite precipitation datasets has shown that TAMSAT performed better on a daily basis in Burkina Faso [40] . For merging the precipitation data from the selected rain gauges with TAMSAT data, bias correction and interpolation methods were employed throughout the Climate Data Tools ( https://iri.columbia.edu/our-expertise/climate/tools/cdt/ ), an R-based software developed by the International Research Institute for Climate and Society (IRI). The empirical quantile mapping method [41] was used to compute the bias correction factor between station observations and TAMSAT estimates at the station location, while the regression kriging [42] was used as the interpolation method aiming to capture the spatial dependence of the residuals. These methods have been extensively used for operational applications [43] [44] .

ETo is a critical parameter for crop water requirement estimation and therefore for the soil water balance of the crop root zone. To estimate ETo, air temperature, relative humidity or dewpoint temperature, solar radiation, wind speed, and elevation data are required [45] . However, in Burkina Faso, only ten weather observation stations (synoptic stations) with long-term data records for these weather parameters are available. To address the specific need of ETo for this study, ERA5 gridded data at a resolution of 0.44˚ × 0.44˚ along with a Digital Elevation Model Data ( https://srtm.csi.cgiar.org/srtmdata ) have been used to calculate ETo, using the FAO-Penman-Monteith Equation (1) following [45] . ERA5 reanalysis at this resolution provides physically consistent meteorological variables for ETo calculations while aligning with the district-level crop yield data available for model validation.

$ETo=\frac{0.408\times \Delta \times \left(R_n-G\right)+\gamma \times \frac{900}{T+273}\times u_2\times \left(e_s-e_a\right)}{\Delta +\gamma \times \left(1+0.32\times u_2\right)}$ (1)

where: R_n is the net radiation flux density on the crop surface (MJ∙m⁻²∙d⁻¹); G is the soil heat flux density (MJ∙m⁻²∙d⁻¹); T is the average daily air temperature (˚C); u₂ is the wind speed at 2 m high (m∙s⁻¹); e_s is the saturation vapor pressure (kPa); e_a is the actual vapor pressure (kPa); Δ is the slope of vapor pressure-temperature curve (kPa∙˚C⁻¹) and γ is the psychometric constant (kPa∙˚C⁻¹).

Soil hydraulic characteristics such as the upper limit of volumetric water holding capacity (FC), the lower limit of water holding capacity (PWD), drainage coefficient (τ), and hydraulic conductivity at saturation (Ksat) are necessary for running AquaCrop. Soil parameters including organic carbon and the fractions of clay, silt, and sand were obtained from the latest version (2.0) of the Harmonized World Soil Database (HWSD) [46] . For a given grid cell at a resolution of 0.44˚ × 0.44˚ in the study area, these retrieved soil parameters were used to calculate textural classes based on the USDA classification, and then summarized to determine the dominant soil texture for each of the soil layers. Computed soil textures are then used to derive soil hydraulic characteristics such as the upper limit of volumetric water-holding capacity (FC), the lower limit of water-holding capacity (PWD), drainage coefficient (τ), and hydraulic conductivity at saturation (Ksat), which are necessary for running AquaCrop [34] . The hydraulic characteristics for the dominant soil texture were based on the FAO database implemented in AquaCrop. In this study, the maximum depth of the soil was set to 1.5m for maize crop rooting, which corresponds to the first six layers in HWSD v2.0.

In crop simulation, the availability of observed crop yield data is crucial to ensure that crop models accurately represent local agricultural conditions. We obtained maize crop production and cultivated areas from the Direction Générale des Etudes et des Statistiques Sectorielles of the Burkina Faso Ministry of Agriculture, Animal Resources and Fisheries. These data are available at the province level, covering the 45 provinces in Burkina Faso, and span the period from 2009 to 2022. We performed crop yield calculations (kg/ha) at a resolution of 0.44˚ × 0.44˚. The yield for a specific grid cell is calculated using a composite weighted average that considers all provinces that share the same grid. The cultivated areas of the relevant provinces were used as the weights in Equation (2).

$Y_{grid}\left(kg\cdot h{a}^{-1}\right)={\displaystyle {\sum }_{i=1}^n{w}_i\times Y_{province\left(i\right)}\left(kg\cdot h{a}^{-1}\right)}$ (2)

where Y_grid (kg/ha) represents the gridded crop yield, Y_province(i) represents the crop yield in province i, w_i represents the fraction of cultivated area of province i within the grid cell, and n represents the number of provinces that share the area of the grid cell.

Calibration is a systematic process of optimizing model parameters to minimize the difference between simulated outputs and observed data, thereby improving model accuracy for specific environmental conditions. For crop models, this process involves adjusting biophysical parameters and management practices to better reflect local farming conditions. In this study, we employed a Genetic Algorithm optimization approach to calibrate selected maize crop and management parameters. The calibration methodology combines mathematical optimization with local agricultural knowledge. Specifically, the parameter selection and boundaries were driven by the specific challenges of Sahelian agriculture and constrained by local maize varieties and well-documented agronomic practices from the Institut de l’Environnement et de Recherches Agricoles (INERA) in Burkina Faso (Olaoye et al., 2009).

The crop parameters were initially drawn from the generic maize crop file in AquaCrop [32] , while the growing period durations were established based on the main maize cultivar categories in use in Burkina Faso: short-duration (75 - 80 days), medium-duration (90 - 100 days), and long-duration (105 - 120 days) varieties [47] [48] . Similarly, plant density parameters computed based on row spacing and plant spacing along rows were established based on comprehensive field surveys on agronomic practices of the main cultivars used in smallholder farming systems across Burkina Faso [49] [50] . As sowing stratey, row spacing was fixed at 0.80 m, while plant spacing along rows varied from ~0.25 m to 0.50 m according to cultivars. The phenological development stages, which determine the length of the growing period, were parameterized according to [32] and [51] . Among management parameters, we prioritized the calibration of relative weed cover and soil fertility stress coefficients, as these factors show high spatial and temporal variability in smallholder farming systems and significantly impact crop water productivity [52] . The following Table 1 summarizes the key parameter set selected for calibration. The calibration approach ensures that our optimized parameters represent not just mathematically optimal solutions but also practically viable recommendations grounded in local agricultural reality.

Genetic algorithms (GAs) are adaptive heuristic search algorithms inspired by the principles of natural selection. They are based on genetic mechanisms such as selection, crossover, and mutation [53] [54] . Unlike traditional search methods, they are particularly suitable for global optimization problems that involve finding the best solution within a large and complex space [55] . However, the performance of GAs is heavily influenced by a number of factors, one of the most pivotal being the objective function used [56] . In this study, the genetic algorithm (GA) is combined with a designed objective function (fobj) to optimize parameter sets for location-specific calibration of maize crops. The calibration process through the objective function performs simulations at a grid cell resolution of 0.44˚ × 0.44˚ using AquaCrop’s open-source Fortran code as a subroutine within the GA. For each grid cell, climate and soil data remain constant across GA iteration steps, while selected crop and management parameters are optimized according to the objective function.

The objective function takes into account two performance metrics, namely the coefficient of determination (R2) and the Relative Absolute Error (RAE). These two statistics are calculated based on the observed yield and the simulated yield using the AquaCrop model. R2, RAE and fobj are expressed as the following Equation (3). To ensure the significance of the Pearson linear correlation (R), the p-value has been set to a maximum of 5%. These statistical measures guide the GA in optimizing location-specific crop and management parameters for maize cultivation across Burkina Faso. Ultimately, the success of a GA requires a thorough understanding of the problem landscape and depends on the careful combination of multiple objective functions to achieve a balanced approach for the optimization tasks [57] .

$R=\frac{{\displaystyle {\sum }_{i=1}^n(O_i-O_{avg})(P_i-P_{avg})}}{\sqrt{{\displaystyle {\sum }_{i=1}^n{(O_i-O_{avg})}^2}{\displaystyle {\sum }_{i=1}^n(P_i-{P_{avg})}^2}}}$

$RAE=\frac{{\displaystyle {\sum }_{i=1}^n\left|O_i-P_i\right|}}{{\displaystyle {\sum }_{i=1}^n\left|O_i-O_{mean}\right|}}$ (3)

$f_{obj}=\{\begin{array}{l}-{\left|\left(RAE\right)\right|}^{\left(1-R^2\right)}\text{if}{R}^2>0.2\text{and}p\text{-value}<5\%\\ -\infty \text{if}{R}^2\le 0.2\text{and}p\text{-value}>5\%\end{array}$

where O_i and P_i represent the observed and simulated yields for the i-th cropping season, respectively. O_avg and P_avg indicate the mean values of observed and simulated yields for the n cropping seasons of the simulation.

In this study, the optimization process aims to maximize f_obj. The highest values of f_obj correspond to those close to zero (dark blue area), as depicted in Figure 4 . Hence, a set of optimal parameter swill yield an RAE close to zero and R² close to one.

The flowchart in Figure 5 depicts the different steps in the process of location-specific crop calibration using a GA. A member of a GA population of candidate solutions is a vector of eight crop and management parameters. For each member, once the growing duration parameter is known through the GA selection process, it is used to derive the comprehensive phenology of the crop. This step completes the user-defined parameters set required to set up the crop file essential for AquaCrop simulation.

Different types and sources of data have been preprocessed to obtain the required input data for AquaCrop runs. From the soil data, the soil fraction of silt, clay, sand, and organic content has been transformed into soil textural classes. The results shown in Figure 6 indicate that there are a total of 7 USDA soil textural classes across Burkina Faso. Sandy loam, sandy clay loam, and clay are the dominant soil textures for all the layers considered, with sandy loam being the most prevalent in the northern half of the country while sandy clay loam and clay dominate the south and southwest regions. The soil depth less than 60 cm is more heterogeneous compared to the layers ranging from 80 cm to 150 cm.

Daily ETo has been computed based on the FAO-Penman-Monteith method and using ERA5 reanalysis data and further re-gridded at a resolution of 0.44˚ × 0.44˚. The results on the long-term mean (period 1981-2020) spatial variation of ETo indicate that it ranges from 5 mm/day to 8 mm/day throughout the country ( Figure 7(a) ). It shows a North-South gradient, with the highest values observed in the northern part and the lowest values in the southern part of the country.

Maize crop yields for the 45 provinces in Burkina Faso from 2009 to 2022 have been re-gridded and used for the calibration process. The average crop yields across the country range from 0.5 t/ha to 2.5 t/ha ( Figure 7(b) ). The lowest yields are located in Northern half of BF, while the highest are found in the southwestern part of Burkina Faso.

The combination of AquaCrop and a Genetic Algorithm (GA) has been implemented in a total of 156 grid-cells. The main objective is to find the global optimum in the search space, thereby maximizing the fitness. The fitness values are calculated through the fitness function which is based on a specific set of parameters heuristically chosen at each step of the GA. Figure 8 illustrates the GA process for a randomly selected set of grid cells. Overall, the results indicate that the global optimum is achieved after at least the 20th generation. The value of the global optimum obtained through the GA process ranges from −0.65 to −0.15 ( Figure 8(b) and Figure 8(c) ), whereas theoretically, the optimal value should be zero, indicating a perfect match between simulated and observed yield. Figure 8(c) demonstrates a scenario of rapid convergence towards the global optimum, while Figure 8(a) and Figure 8(d) depict slower convergence. Figure 8(b) represents a hybrid case with slow convergence until the 60th generation, followed by rapid convergence.

For a given grid cell, the coefficient of determination ( Figure 9(a) ) and the relative absolute error ( Figure 9(b) ) were calculated using pairwise simulated and observed yield data for the period 2009-2022. The simulated yield was performed using AquaCrop in combination with the optimal crop and management location-specific parameter sets obtained throughout the GA crop calibration processes. Our findings reveal that the coefficient of determination (R2) varies from 0.20 to 0.65, whereas the yield deviation (RAE) varied from 8% to 36% across the country. Notably, the southwestern part of the country exhibits the highest R2 values, while the northern part exhibits the lowest values. Among the 156 grid cells, approximately 60% show yield deviations within the range of 21% to 30%.

Crop and management parameters derived from the GA process of location-specific calibration for maize cropping have been analyzed for the three climatic zones in BF. The results show that regardless of the climatic zone, a base temperature of 10˚C and an upper temperature of 32˚C are required for at least 50% of grid cells ( Figure 10(a) and Figure 10(b) ). The duration of the growing season ( Figure 10(c) ) and the harvest index ( Figure 10(f) ) vary significantly among the three climatic zones, with the highest values found in the Soudanian zone and the lowest values in the Sahelian zone. The mean plant sowing density ( Figure 10(d) ) is approximately 31,000 plants per hectare, but there is considerable variability within each climatic zone. The flowering duration ( Figure 10(j) ) is consistent across the three climatic zones, lasting about 2 to 3 weeks. Crop development stages ( Figure 10(g) , Figure 10(h) and Figure 10(i) ) show a similar trend in variability as the duration of the growing season. In terms of management strategies, the overall fertility rate parameters range from 20% to 55%, with mean values of 35%, 39%, and 49% for the Sahelian, Soudano-sahelian, and Soudanian zones, respectively ( Figure 10(e) ). The mean weed cover ( Figure 10(e) ) is approximately 36%, with the Sahelian and Soudano-sahelian zones showing high variability and the Soudanian zone showing low variability.