The Capabilities of WRF in Simulating Extreme Rainfall over the Mahalapye District of Botswana

Abstract

Flooding episodes caused by a heavy rainfall event have become more frequent, especially during the rainfall season in Botswana, which poses some socio-economic and environmental risks. This study investigates the capability of the Weather Research and Forecasting (WRF) model in simulating a heavy rainfall event that occurred on 26 December 2023 in Mahalapye District, Botswana. This event is one among many that have negatively impacted the lives and infrastructures in Botswana. The WRF model was configured using the tropical-suite physics schemes, i.e., (Rapid Radiative Transfer Model, Yonsei University planetary boundary layer scheme, Unified Noah land surface model, New Tiedtke, Weather Research and Forecasting Single-Moment six-class) on a two-way nested domain (9 km and 3 km grid spacing) and was initialized with the GFS dataset. Gauged station data was used for verification alongside synoptic charts generated using ECMWF ERA5 dataset. The results show that the WRF model simulation using the tropical-suite physics schemes is able to reproduce the spatial and temporal patterns of the observed rainfall but with some notable biases. Performance metrics, including RMSE, correlation coefficient, and KGE, showed moderate to good agreement, highlighting the model’s sensitivity to physical parameterization and resolution. The results of this study conclude that the WRF model demonstrates promising potential in forecasting extreme rainfall events in Botswana, but more sensitivity tests to different parameterization schemes are needed in order to integrate the model into the early warning systems to enhance disaster preparedness and response.

1. Introduction

Global concerns over extreme events are highlighted by [1], Indicating that several extreme climate events are increasing in frequency and intensity due to anthropogenic climate change, hence raising the likelihood of impacts associated with urbanization. Ref. [2] also stated that climate change could cause increased frequency in droughts and floods. Refs. [1,2] observed that climate change may cause climate variability, changes in average precipitation, and extreme events occurrences. Ref. [3] provided an analysis of flood occurrences and their impacts over a five-year period (2015–2019). The report indicated that flood episodes have increased in Botswana and are caused by heavy rainfall, inadequate drainage systems, and growth of settlements into locations susceptible to flooding. Ref. [3] showed that the most affected areas in Botswana were Ghanzi, Northeast, Central, Kweneng, Kgatleng, and Southern Districts, where infrastructure and human settlements have been significantly impacted. The findings from [3] indicated an increase in the occurrence and severity of floods in Botswana, with a recommendation that integrating meteorological forecasting, hydrological modeling, and community-based disaster preparedness will be essential in mitigating the socio-economic impacts of future flood events. With increasing occurrences of extreme events, especially floods, their impacts leave lasting effects on affected communities. With recurring floods, there are significant impacts on both socio-economic and environmental impacts. The study by [4] highlighted that the 2016/2017 floods caused damage to property and infrastructure, livestock losses, crop damage, and human health impacts. The authors estimated that the financial cost of the damage was around BWP 2,407,436, and this amount excludes crop losses since there were no monetary estimates done from the study. Ref. [5] did a study on the floods that affected the Okavango Delta communities in 2004 and 2009. The researchers found that the floods caused crop losses in Tubu and Shorobe Villages of 67% and 72%, respectively. Ref. [6] assessed rainfall variability and vulnerability to floods and droughts. The study highlights that Mahalapye’s flood vulnerability is increased by the presence of two rivers (the Mahalapye River and the Rakabeswa River), which pass through the town and settlement areas of Mahalapye.

Intense precipitation accompanied by strong winds often highlights the presence of severe weather systems. A significant component of what makes weather patterns develop is the continuous transfer of energy, moisture, and momentum between the terrestrial surface and the atmosphere. Meteorologists aim to figure out how these relationships work because they are key to understanding why heavy rain takes place [7,8]. Global Circulation Models (GCMs) are used to make weather forecasts over a range of time periods [9,10]. GCMs are not able to capture the mesoscale processes that contribute to rainfall. They do not show important mesoscale and local sub-grid features like convective processes (updrafts and downdrafts associated with thunderstorms and other convective systems), and they tend to overestimate rainfall [10,11]. To improve the representation of regional and local weather-related phenomena, Limited Area Models (LAMs) were developed as a dynamical down-scaling approach [12,13]. Unlike GCMs, LAMs operate at higher grid spacing (1–50 km), allowing for better simulation of orographic precipitation, convective processes, and land–sea interactions [14,15].

LAMs serve as a supplementary research approach, facilitating further process investigations and the simulation of regional and local situations using the downscaling technique [16]. The downscaling technique using LAMs magnifies the outputs of GCMs. LAMs can generate improved regional meteorological data that aligns with the large-scale circulation from global reanalysis data or general circulation models (GCMs), utilizing precise representations of physical processes and high grid spacing to capture complex topography, land–sea contrasts, and land use [17]. Researchers, including [18], have been developing a LAM called the WRF model—an advanced NWP system designed for both operational forecasting and atmospheric research. WRF is an open-source community model used for research and numerical weather prediction [18], as well as medium-range forecasting. The WRF model was developed by the National Center for Atmospheric Research (NCAR); however, the Mesoscale and Microscale Meteorology (MMM) division of the University Cooperation for Atmospheric Research (UCAR) continues to maintain and update the WRF model [10,19]. In the late 1990s, WRF developers began to produce a research-operations system for next-generation NWP capacity. The WRF model is a completely compressible, non-hydrostatic mesoscale NWP model with topography following sigma coordinates [10,19]. Ref. [20] applied the WRF model at a high grid-spacing of 2.5 km to study extreme weather events and integrate the model into the Extreme Rainfall Detection System (ERDS). The authors found out that it can improve early detection and warning of extreme rainfall events, especially by assimilating observations into the WRF model.

Between 16 and 23 January 2013, a total of 4210 people (at least 842 families) suffered as their households were destroyed by widespread thunderstorms, which caused extensive flooding in the Central District of Botswana, also affecting the Mahalapye District [21]. The floods displaced many residents and resulted in the majority of them living in tents for weeks, even while their houses were being rebuilt. This episode highlighted, among other things, the absence of a reliable early warning and prediction system. A solution to this problem might be NWP, specifically, dynamic modeling. The low grid-spacing of most GCMs is unsuitable for heavy rainfall analysis; thus, LAMs can be utilized to estimate rainfall at small grid spacing [22]. This study aims to assess the ability of the WRF model to simulate high rainfall events within the Mahalapye District of Botswana, highlighting the limitations and areas for improvement in the WRF model, and it adopts the WRF tropical suite physics scheme configuration, which is recommended for the tropical belt and subtropical regions [23]. This suite has been tested and consistently demonstrates good performance under tropical thermodynamic conditions. Ref. [24] used the tropical suite as a baseline and later applied sensitivity tests, and it showed that variations in microphysics (WSM6 vs. Thompson) and PBL formulations (YSU, MYJ, ACM2) exerted more influence on rainfall structure localization. Their results showed that the single-moment WSM6 scheme outperformed the double-moment Thompson scheme while non-local PBL formulations (YSU, ACM2) were found to produce realistic low-level mixing and delayed convective triggering compared with the local WYJ. Also, ref. [25] reported that in the southern Africa region, WSM6 coupled with non-local PBL schemes best represented convective intensity and timing, highlighting that PBL dynamics plays a big role over rainfall structure than microphysical complexity. Ref. [26] also confirmed that RRTM-YSU-WSM6-Noah configuration (tropical suite) was able to reproduce the diurnal evolution of boundary-layer fluxes, cloud depth, and storm organization, which outperformed a different scheme used in mid-latitudes called CONUS suite. The choice of the tropical suite for this study follows the collective studies mentioned, which all concluded that the tropical suite configuration is relevant for the southern Africa region.

Even though there have been publications on extreme rainfall and flooding in Botswana, there remains a limited understanding of the capability of high-resolution NWP prediction models to realistically reproduce the physical structure, intensity, and spatial distribution of localized heavy rainfall events. Most WRF studies conducted in Botswana have focused on investigating the tropical cyclones that made landfall over the continent and subsequently affected Botswana as a tropical depression. Ref. [27] investigated tropical cyclone Dineo using WRF and found that different model configurations produced generally similar rainfall simulations. However, observations still remain relevant in modern NWP. This highlights the gap and motivates the application of the WRF model to simulate extreme rainfall events in Botswana that are not associated with tropical cyclones.

The WRF model has not been systematically evaluated for non-tropical cyclone extreme rainfall events, specifically in Botswana, such as the Mahalapye District. This study, therefore, serves as a foundational assessment of WRF performance in this context, contributing to the WRF modeling community while providing a region-specific insight for operational forecasting. The results are intended to inform subsequent model development efforts, including targeted model physics sensitivity experiments, and instances where deficiencies are identified; the application of WRF data assimilation techniques is used to enhance initial conditions and improve predictive skill for extreme rainfall events over Botswana.

2. Materials and Methods

2.1. Study Area

Botswana is a landlocked country in southern Africa, located between latitudes 18–27° S and longitudes 20–29° E. The country spans 582,000 km² and has a population of over 2 million [28]. Botswana has a new addition of administrative boundaries, which were initially 10 in total. The districts are Ngamiland, Chobe, Gantsi, Kgalagadi, Southern, Southeast, Kgatleng, Kweneng, Central and Northeast. Botswana has divided its districts by subdividing some of the existing districts, bringing the total to twenty-six (26) districts. The Central District has been divided into eight (8) more districts. The Mahalapye District is the primary focus area for this study and is geographically located in the southeastern parts of the Central District (see Figure 1 below). The Mahalapye District population is approximately 48,431 according to Statistics Botswana [29].

2.2. Source of Data

The Global Forecasting System (GFS) model provides global coverage forecasts across the globe with a horizontal resolution of approximately $0.25°\times 0.25°$, and this is equivalent 28 km grid spacing at the equator. The data is freely available from the National Centers for Environmental Prediction (NCEP) file transfer server (FTP) https://ftp.ncep.noaa.gov/data/nccf/com/gfs/prod/, (accessed on 1 January 2024). The dataset is updated daily and includes parameters such as temperature, humidity, wind speed and direction, geopotential height, surface pressure, and precipitation. The data is provided in GRIB2 format and updated every six hours (00, 06, 12 and 18 UTC forecast cycle) and available in three hours intervals for forecast periods of up to 16 days ahead [30].

For this study, the WRF model used GFS datasets, from the 00 UTC cycle as initial and lateral boundary conditions. This dataset contains GFS model outputs at a global scale detailing the atmospheric state according to GFS and provides WRF model initialization from short to medium range forecasting, and it can be used for extreme weather event model simulations, as highlighted by [31,32]. GFS data has also been used in Botswana by [27] and also in upper Ganga Basin by [22]. BDMS has 18 synoptic stations and over 600 rainfall stations scattered across the country. The rainfall stations are placed in most of the government owned premises such as schools, police stations, and some army bases, to mention a few. For this study, a total of six (6) stations available in the Mahalapye District were used, namely, Kalamare Primary School, Mahalapye Meteorological Station, Machaneng Police, Pallaroad Kgotla, Mookane Primary School and Tirelo Primary School, see (Figure 1). Their respective latitudes, longitudes, and altitude are shown in

Table 1. Station names used in the Mahalapye District.

Station Name	Lat	Lon	Altitude (m)
Mahalapye Meteorological Station	$23.1166°$ S	$26.8333°$ E	1017
Mookane Primary School	$23.6841°$ S	$26.6599°$ E	951
Tirelo Primary School	$23.7481°$ S	$26.4717°$ E	957
Kalamare Primary School	$22.9333°$ S	$26.5666°$ E	1074
Machaneng Police	$23.1886°$ S	$27.4892°$ E	888
Pallaroad Kgotla	$23.4025°$ S	$26.6968°$ E	1002

below.

The data used was supplied by the BDMS, which maintains a quality standard and ensures that the gauged weather stations are serviced on an annual basis by a team of engineers before the onset of a rainfall season. The data is readily available from BDMS and can be used for research purposes as it has been used by [33] who employed monthly precipitation data from 60 stations provided by BDMS, covering the data period from 1981 to 2016, to analyse rainfall patterns in Botswana; additionally, ref. [34] utilized daily precipitation data from 76 rain gauge stations across Botswana to develop regional Intensity Duration Frequency (IDF) curves for ungauged sites, aiding in the estimation of design storm depths. Ref. [35] analyzed historical monthly rainfall data from the Palapye police weather station, provided by BDMS, to assess climate change and variability in the region and lastly ref. [36] explored the influence of climate variability on rainfall characteristics in the Shashe catchment, utilizing observed data from 1981 to 2020 provided by BDMS and used Coupled Model Intercomparison Project Phase 5 General Circulation Models to produce future projections of rainfall for 2021–2050.

To complement the gauged rainfall observations from the BDMS network, large-scale atmospheric fields were derived from the European Center for Medium-Range Weather Forecast (ECMWF) ERA5 reanalysis. The ERA5 dataset provides a consistent representation of the state of the global atmosphere, generated through an advanced four-dimensional variational (4D-Var) data assimilation system that also integrates satellite, surface, and upper-air observations [37]. The horizontal resolution of ERA5 data is $0.25°\times 0.25°$ and a total of 137 hybrid sigma-pressure levels, which extends from the surface to around 0.01 hPa, which offers detailed diagnostics of vertical atmospheric structure. The data is also suitable for assessing the large-scale drivers and moisture transport process that influences convection development over southern Africa [38,39]. In addition, the ERA5 dataset has shown strong agreement with in situ and satellite observations across Africa when compared to the earlier ERA-interim dataset [40], especially key meteorological variables such as precipitation and humidity [41].

ERA5 reanalysis data were used to characterize the synoptic-scale atmospheric conditions associated with the extreme rainfall event in the Mahalapye District. ERA5 integrates a wide range of observations and remotely sensed and provides a dynamically consistent representation of the observed atmosphere [37]. In contrast, GFS analyses were used as initial and lateral boundary conditions for the WRF simulations in order to replicate an operational NWP framework. Using ERA5 for synoptic diagnosis and GFS for model forcing allows independent evaluation of WRF performance while avoiding the circularity that would arise from using the same dataset for both atmospheric characterization and model initialization.

2.3. Numerical Study Design

The current study uses the Advanced WRF model, version 4.4.1, released in August 2021. The WRF version 4.4.1 was an update of the WRF version 4.4. The WRF model produces output every three hours using a two-way nesting in this simulation. In a two-way nesting configuration, the lateral boundary conditions are sent from the parent domain to the child nest, and the values in the child domain (higher resolution domain) are averaged to the corresponding value in the parent domain (lower resolution domain). The outer domain (parent domain), which has a horizontal resolution of 9 km, covers most of southern Africa and parts of the Atlantic and Indian Oceans. The 9 km domain setup is nested with an inner domain (Child domain) at a horizontal resolution of 3 km centered over Botswana, covering some parts of its neighboring countries, as shown in Figure 2.

The WRF model was initialized with GFS data, which has a grid spacing of approximately 28 km (0.25°). GFS pgrb2 format is used to provide both initial boundary conditions and lateral boundary conditions with a time step of 3 h. The model used the Runge–Kutta three-time integration scheme, which solves equations representing atmospheric processes over small time intervals to produce a forecast [18]. The WRF model simulation was initialized on 24 December 2023 at 00:00 UTC with a lead time of 48 h to simulate the extreme event. In order to account for the 48 h of severe rain that occurred on 26 December 2023, the model is set up to produce a forecast for a maximum of 120 h or five days of simulation.

Table 2. WRF model configuration for simulation of a heavy rainfall event.

Model Options	Specifications
Model type	Non-hydrostatic
Domains	two (2) (with a two-way nested domains)
Grid resolution (spacing)	parent domain 1: 9 km × 9 km; (312 × 301 grid)
	nested domain 2: 3 km × 3 km; (403 × 412 grid)
Map projection	Mercator
Initial conditions	NCEP GFS (3-h interval)
PBL Schemes	YSU scheme
Cumulus schemes	A newer Tiedtke scheme
Microphysics schemes	WSM 6-class graupel scheme
Shortwave radiation scheme	RRTMG scheme
Longwave radiation scheme	RRTMG scheme
Land surface model	Unified Noah land surface model

shows the WRF model tropical suite configuration. The model simulation used the tropical suite configuration since Botswana is located within the sub-tropics, and the “tropical suite”, as proposed by [42], integrates several schemes that are suitable for tropical environments (see

Table 3. WRF Tropical Suite physics schemes.

Scheme	Description
RRTM	Rapid Radiative Transfer Model used to improve radiative transfer calculations during the WRF run; developed by [43].
YSU	Yonsei University planetary boundary layer scheme introduced by [44]; represents boundary layer processes.
Noah LSM	Unified Noah land surface model used to capture land surface interactions; developed by [45].
New Tiedtke	Updated cumulus scheme by [46], used in cumulus convection representation.
WSM6	Weather Research and Forecasting Single-Moment 6-class microphysics scheme developed by [44]; provides detailed microphysical processes.

below).

2.4. Model Evaluation Statistics

In assessing the accuracy of the WRF model simulation against observed rainfall data, bias, correlation, and event probability of detection capabilities were used, with a key focus on extreme rainfall. This study used a range of widely accepted statistical metrics, i.e., Root Mean Square Error (RMSE), Kling–Gupta Efficiency (KGE), Percentage Bias (% Bias), Pearson’s Correlation Coefficient (r), Probability of Detection (POD) and Variability Ratio (VR).

As demonstrated by [47], Equation (1), the RMSE, shows the average magnitude of the error between the observed rainfall from BDMS gauged stations and the corresponding WRF model—simulated rainfall. The method is effective in identifying extreme events since it imposes more penalties on large errors [48]. A lower RMSE value translates to enhanced model accuracy, while a higher value indicates a lower model accuracy.

Percent Bias (PBIAS) is a statistical metric used to show how frequently a model tends to overestimate or underestimate observations and computes it as a percentage. Ref. [49] highlighted that PBIAS is an important measure for hydrological, meteorological and atmospheric modeling for identifying systematic errors in simulations. PBIAS is calculated as shown in Equation (2).

Pearson’s correlation coefficient (PCC) is mainly used for model verification because the metric has the ability to quantify how well a model is able to reproduce observed values [50,51]. In the context of evaluating the WRF outputs, the metric is able to measure the variation between the simulated and the observed rainfall, and as a result, it assesses how well the WRF model captures the temporal and spatial patterns in observed rainfall. PCC is calculated as follows: Equation (3).

Probability of Detection (POD), Equation (4), is defined as the ratio of the model being able to predict/reproduce an occurrence (hits) to the total number of observed events (hits + misses) [51].

Lastly, the Kling–Gupta Efficiency (KGE) was used to address some limitations of the traditional Nash–Sutcliffe Efficiency by ensuring a more balanced representation of error sources. The KGE is illustrated in Equation (5). KGE values range from $-\infty$ to 1, with 1 indicating an agreement between the observations and the model prediction, while 0 suggests that the model does not perform better than the mean observed data and negative values indicate poor performance [52]:

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{(P_i-O_i)}^2}$$

(1)

$$\mathrm{PBIAS}=100\times {\displaystyle \frac{{\sum }_{i=1}^n(P_i-O_i)}{{\sum }_{i=1}^n{O}_i}}$$

(2)

$$r={\displaystyle \frac{{\sum }_{i=1}^n(P_i-\overline{P})(O_i-\overline{O})}{\sqrt{{\sum }_{i=1}^n{(P_i-\overline{P})}^2}\sqrt{{\sum }_{i=1}^n{(O_i-\overline{O})}^2}}}$$

(3)

where

• r = Pearson’s correlation coefficient;
• $P_i$ = WRF-simulated rainfall value;
• $O_i$ = observed rainfall value (from BDMS gauges);
• $\overline{P},\overline{O}$ = mean of simulated and observed rainfall, respectively;
• n = total number of observations (stations or time steps).

$$\mathrm{POD}={\displaystyle \frac{\mathrm{Hits}}{\mathrm{Hits}+\mathrm{Misses}}}$$

(4)

where

• Hits = Number of times the event was correctly predicted;
• Misses = Number of times the event occurred but was not predicted.

$$\mathrm{KGE}=1-\sqrt{{(r-1)}^2+{(\beta -1)}^2+{(\gamma -1)}^2}$$

(5)

Haversine Formula

Many model evaluation studies often interpolate simulated model outputs to observational locations, which implies comparing gridded data and station observations. A commonly used approach is selecting the four nearest model grid points surrounding the gauged station location (Equation (6)) and applying a weighted average Equation (7) to estimate the model value at the station’s exact location [53,54]. To apply the weighted average estimation to the gridded data, the Haversine formula is often used to calculate the distances closest to the station coordinate and is expressed as follows:

$$d=2rarcsin\left(\sqrt{{sin}^2\left({\displaystyle \frac{\Delta \varphi }{2}}\right)+cos{\varphi }_{1}cos{\varphi }_2{sin}^2\left({\displaystyle \frac{\Delta \lambda }{2}}\right)}\right)$$

(6)

where

• d = great-circle distance between two points (km);
• r = Earth’s mean radius (approximately 6371 km);
• ${\varphi }_1,{\varphi }_2$ = latitudes of the two points (in radians);
• ${\lambda }_1,{\lambda }_2$ = longitudes of the two points (in radians);
• $\Delta \varphi ={\varphi }_2-{\varphi }_1$;
• $\Delta \lambda ={\lambda }_2-{\lambda }_1$.

After identifying points closest to the station, the interpolation of model outputs to station locations was performed using the Inverse Distance Weighting (IDW) method. In this approach, weights are assigned inversely proportional to the distance between each grid point and the station such that the closest grid point to the gauged station will receive the highest weight, and the interpolated value is derived from a weighted mean of the four surrounding grid values [55]. This method has been used by [10,53,54]. This method may face some challenges in regions with complex terrain, implying that atmospheric conditions vary significantly over a short horizontal distance. In cases where topography is a bit complex, distance-based interpolation alone may not fully capture the representativeness of the model values at station locations since variables like precipitation and temperature are sensitive to topography. The terrain of Botswana is generally flat with less complex terrain; hence, this method has been adopted for this study.

$$P\left(\mathrm{station}\right)={\displaystyle \frac{{\displaystyle \sum _{k=1}^n{w}_k{P}_k}}{{\displaystyle \sum _{k=1}^n{w}_k}}},w_k={\displaystyle \frac{1}{d_k^2}}$$

(7)

where

• $P\left(\mathrm{station}\right)$ = interpolated precipitation at the station (mm);
• $P_k$ = precipitation value at the $k\mathrm{th}$ model grid point (mm);
• $w_k$ = weight assigned to the $k\mathrm{th}$ grid point;
• $d_k$ = haversine distance between the station and the $k\mathrm{th}$ grid point (km);
• n = number of nearest grid points used ($n=4$).

3. Results and Discussion

The results of this study are organized into four subsections. Section 3.1 explores the influence of topography on WRF simulations, showing how terrain representation contributes to storm development and rainfall distribution. Section 3.2 outlines the synoptic conditions that played a role in the extreme rainfall event which occurred in the Mahalapye District. Section 3.3 displays the observed precipitation from gauged stations, which serves as a reference against which the WRF output performance is assessed. Section 3.4 evaluates the WRF simulations by comparing simulated precipitation and ERA5 outputs with observed data and applying statistical metrics to produce model skill.

3.1. Interaction of Topography with Synoptic Systems

For this research, the Shuttle Radar Topography Mission (SRTM) DEM was used to provide near-global elevation coverage at a horizontal resolution of approximately 1 km. The SRTM data was derived from radar interferometry during the Space Shuttle mission in February 2000 [56]. SRTM has established itself as a standard reference in topographic and hydrometeorological research throughout Africa for purposes including terrain-based rainfall analysis, flood modeling, and land surface characterization [57,58]. The SRTM dataset applied in this study serves as a control representation of the actual landscape against which the coarser resolution ERA5 and dynamically downscaled WRF topographies (see Figure 3 below) were compared to assess how well the models resolve real-world elevation gradients over Botswana.

In ERA5, the coarse terrain resolution (approximately 31 km) results in a smoothed elevation field that fails to capture the topographic gradients and steep slopes surrounding the Mahalapye area. This implies that the associated low-level convergence and upslope lifting have been underestimated, leading to a weaker rainfall signal in ERA5. Compared to ERA5, the WRF simulations at 9 km (d01) and 3 km (d02) capture complex topographic cliff sides and small-scale ridges across the central parts of Botswana. With enhanced terrain representation, the WRF model simulations are improved, especially the local wind channelling and moisture advection from the east, which enables more realistic convective developments and anchoring along the elevated zones of the Mahalapye District.

3.2. Synoptic-Scale Conditions Prevailing on the Day of the Event: 26 December 2023

Figure 4 shows the mean sea level pressure (MSLP) associated with 26 December 2023 extreme rainfall event which were downloaded from South African Weather Services (SAWS) and ECMWF respectively. The ECMWF ERA5 reanalysis datset was used to produce the bottom image. The two analyses are in agreement as they show that at the surface, the mean sea level pressure (MSLP) indicated a surface low-pressure system, which aided in the convergence of moist air from the Indian/Atlantic Oceans towards the land, leading to increased atmospheric instability. The surface charts also show that the pressure gradient on the day was slightly steep, which may suggest the likelihood of strong winds. On the day of the extreme weather event, both heavy rainfall and strong winds were experienced, and this further supports the interpretation that the conditions experienced at the surface played a significant role in the extreme weather event observed.

The images in Figure 5 below were generated using ECMWF ERA5 dataset. (top image) shows negative vorticity or medium-level instability over the northeastern and central parts of Botswana. Vorticity is associated with the spin or rotation in the atmosphere, and it often facilitates upward motion, which can lead to cloud formation and intensification. This phenomenon is highlighted by the ERA5 map and often leads to the development of precipitation. Both charts in Figure 5 show a trough at the 500 hPa level extending across southern Africa.

Such features are known to cause atmospheric instability at a medium level, increasing the likelihood of storms and enhancing rainfall potential. Figure 5 (bottom) shows precipitation recreated using GFS data from website https://noaa-gfs-bdp-pds.s3.amazonaws.com/gfs.20231225/00/atmos/gfs.t00z.pgrb2.0p25.f012 and https://noaa-gfs-bdp-pds.s3.amazonaws.com/gfs.20231225/00/atmos/gfs.t00z.pgrb2.0p25.f036 (accessed on 22 December 2025), which does indicate that indeed precipitation was forecasted to occur at Mahalapye District. It should also be noted that prior to 26 December 2023, there were previous days within the same area of Mahalapye District where rainfall was recorded, which may have contributed to soil water saturation, and this might have increased the occurrence of the flash flood that was experienced on the day.

3.3. Observed Precipitation from DMS

BDMS has a sparse network of Automatic Weather Stations (AWS) across the country. At the time of this publication, BDMS has a total of twenty (20) operational AWSs; for this reason, the rainfall observations used are derived from conventional rain gauge stations, which have a denser spatial coverage compared to the AWS network. These gauge stations provide reliable 24 h accumulated rainfall totals but do not resolve sub-daily or hourly rainfall evolution. As a result, the model evaluation focuses on daily rainfall total and their spatial distribution rather than on the diurnal timing of rainfall. Figure 6 below shows the observed rainfall data from the network of DMS rain gauges at DMS. The data shows that rainfall was recorded for 26 December 2023. The highest amount of rainfall received was at Mahalapye Meteorological station (114 mm), and the lowest rainfall was received at Kalamare Primary School (24 mm) in Mahalapye District.

3.4. WRF Evaluation

Relative humidity (RH) and wind patterns at both 800 hPa and 500 hPa (Figure 7) show the moisture advection and mid-tropospheric dynamics during 26 December 2023, in which the Mahalapye District experienced heavy rainfalls resulting in a flooding episode. At 800 hPa, as shown in Figure 7, ERA5 data shows that the eastern side of Botswana experienced an enhanced humidity exceeding 80%, which was also visible in the surrounding countries (Zambia, Zimbabwe, Mozambique, and some eastern parts of South Africa), with a strong indication of moisture transport from the Indian Ocean. The inflow of moisture was driven by low-level easterly flow, which is mostly associated with an anticyclonic motion (high-pressure) system ridging in from the Indian Ocean into the southeast of southern Africa, a configuration consistent with the moisture conveyor described by [38] during a squall-line development over the South African Highveld. The outputs from WRF at both the 9 km and 3 km domains slightly reduced the moisture inflow but still maintained saturated pockets of moisture in the northeastern parts of the country. Both ERA5 and WRF at 800 hPa suggest that the model adequately captured near-surface moisture convergence, which plays a significant role in the deep convection initiation.

At 500 hPa, both ERA5 and WRF display a dry mid-troposphere over some parts of Botswana, including Mahalapye District, but a moist region (RH > 60%) in the extreme eastern parts of Botswana. The vertical contrast in humidity often creates an unstable layer conducive to vigorous convective updrafts similar to mid-level disturbances and dry-air intrusions reported by [39,59,60]. These findings from both the ERA5 dataset and the WRF simulation (d01 and d02) indicate that the extreme rainfall event over Mahalapye District was driven by a mid-level westerly disturbance, which played a significant role in convergence and lifting over the eastern side of Botswana. Refs. [19,38,61] have recognized that such thermodynamic and dynamic interactions are a key precursor for heavy rainfall and mesoscale convection development in the subtropical regions in southern Africa.

It is also equally important to quantify how well the WRF model reproduces RH relative to ERA5 reanalysis data. Figure 8 below is the RH difference between the WRF model simulation at both 9 km and 3 km against ERA5 reanalysis ($\mathrm{WRF}-\mathrm{ERA}5$). Positive indicates overestimation and negative indicates underestimation by the WRF model.

At 800 hPa (top left and top right), the d01 (9 km) WRF model simulation shows significant underestimation of RH biases over the Mahalapye District. The same is observed in the d02 (3 km) WRF model simulation, which shows a more pronounced (slightly saturated color) underestimation, as indicated by the negative RH differences.

At 500 hPa (bottom left and bottom right), both WRF model simulations exhibit a mixture of largely overestimation and near-neutral RH over parts of the Mahalapye District. Having similar results at both d01 and d02 may have been caused by the use of a two-way nesting with feedback enabled, which allows continuous interaction between d01 (9 km) and d02 (3 km). This configuration can lead to moisture adjustments in both domains, leading to largely similar mid-tropospheric and near-surface humidity structures. Future studies will explore if a one-way nesting configuration has contrasting results, and we will be able to see domain-specific behaviors where the parent domains drive the nested domain without receiving feedback from the nested domain; this will allow a clearer assessment of resolution-dependent moisture biases.

The daily total rainfall estimates from ERA5 reanalysis and WRF model (Figure 9) show distinct differences in spatial distribution and intensity across southern Africa and specifically the Mahalapye District during the extreme rainfall event period of 26 December 2023. ERA5 reanalysis data indicates that there was moderate rainfall in Mahalapye District (daily maximum below 20 mm). This may be highly attributed to the course resolution from ERA5 (approximately 31 km), which tends to underestimate localized convection and extreme rainfall magnitudes since it relies on parameterized convection and grid-scale averaging [37,40]. The ERA5 dataset is primarily good at capturing the synoptic scale forcing, which is associated with moisture inflow from the Indian Ocean, but at times, ERA5 outputs fail to explicitly resolve convective updrafts and storm structures at the mesoscale level, as outlined by [19,61].

In contrast, the WRF simulations at 9 km and 3 km resolutions produced a more improved representation of daily total rainfall during the period beginning on 26 December 2023. The 9 km WRF simulation shows an increased total rainfall with a maximum not exceeding 100 mm for the gauged stations in Mahalapye District. At the convection-permitting scale (3 km), the WRF model can explicitly resolve convective processes without the need for parameterization, but for this case, the high-resolution (3 km, d02) WRF simulation had reduced accuracy compared to the 9 km WRF setup.

A list of all observed rainfall events in Mahalapye District, obtained from BDMS against the WRF model run, is shown in (Figure 10) below. The WRF model rainfall events were interpolated using the Haversine and weighted average formula at each station over the Mahalapye District. The observations made from Pallaroad Kgotla and Mahalapye Meteorological Station were poorly simulated by WRF at both 9 and 3 km, respectively (Figure 10). In theory, running the WRF model at 3 km offers a finer grid spacing, and the expectation is to have improved accuracy in the model output. However, in this study, the 9 km WRF model run performed slightly better than the 3 km WRF model run for Mahalapye District. The observations from both Figure 9 and Figure 10 does not necessarily imply that coarser resolution is superior; rather, it may reflect the combined influence of model physics parameterization choices, boundary conditions, and the limited number and spatial distribution of gauged stations used for verification in this study. It has also been shown in [10] that high-resolution modeling produces better spatial distribution of meteorological variables, including rainfall, but does not improve the statistics. Appendix A Figure A1 shows total rainfall in Botswana on 26 December 2023.

Figure 11 below shows the Skew-T diagrams that were derived from WRF model simulations, depicting vertical thermodynamic structure over the Mahalapye District on 26 December 2023 at around 14:00 UTC. The Skew T diagrams show pre-convective environment analysis from the gauged stations under investigation. In both d01 and d02, temperature (in red) and dew point (in green) profiles show a well-mixed boundary layer with a weak inversion near 850–700 hPa and strong drying at 600 hPa. The structures displayed in (Figure 11) are observed more often in continental tropical areas where mid-level subsidence and dry-air mixing have a big effect on convective formation. The features in Figure 11 are consistent with studies by [24,62], who found that the southern Africa convection frequency develops beneath dry mid-tropospheric layers that control storm depth and lifetime.

At all the gauged station sites, surface temperatures exceeded 28–32 °C, while dew point remained near 20–22 °C. This led to low-level dew point depressions of 3–5 °C and a Lifted Condensation Level (LCL) between 712 hPa and around 850 hPa (roughly 1.3 to 2 km). The low LCL heights suggest high boundary-layer humidity and efficient cloud base formation, especially over Machaneng, Mahalapye Meteorological station, and Pallaroad Kgotla, where moisture advection from the northeast was the strongest. Above 700 hPa, dew-point lines diverge sharply from the temperature lines and show a well-defined dry layer that acts as a mixing barrier. This phenomenon of mid-level dryness tends to promote evaporative cooling in downdrafts, highlighted by [62,63], which are common in Botswana’s convective storms. Figure 11 concludes that the WRF model is in a position to show the vertical thermodynamics that contributed to the rainfall activity.

An interesting finding from the 9 km domain is that it slightly outperformed the 3 km model run in terms of statistics, including RMSE, correlation, and the Kling–Gupta Efficiency (

Table 4. Performance metrics for WRF simulations on 26 December 2023 over Mahalapye District.

Model	RMSE (mm)	PBIAS (%)	KGE	r	POD (1 mm)
WRF_3 km	43.529	−56.10	−0.356	−0.038	1.000
WRF_9 km	41.197	−48.22	−0.225	0.020	1.000

). The rule of thumb is that higher-resolution simulations should yield better accuracy since the model will be able to resolve mesoscale processes [9,14]. Running WRF at 3 km recorded a PBIAS of −56.1%, while a 9 km run showed −48.2%, with both exhibiting negative KGE values (−0.356 and −0.225, respectively) and Pearson’s correlation coefficient (−0.038 and 0.020) close to zero to negative correlation with observed values. These findings confirm that WRF struggled with the intensity of rainfall simulation even under increased spatial resolution but managed to correctly capture the timing of rainfall events. The results from this study are supported by previous findings that resolution alone does not always guarantee improvement if model physics are not optimally configured [19,27]. Similar results were also reported by [22] in the upper Ganda Basin, where WRF’s higher-resolution run improved the timing of convection but did not enhance rainfall intensity estimates when evaluated against the eighteen (18) rain gauge stations in their study. For the case of Botswana, ref. [27] found that the ability of WRF to simulate ex-tropical cyclone Dineo was dependent on the microphysics choices, with less sensitivity to model horizontal resolution. This strongly suggests that when it comes to local extreme events, studies should focus on model parameterization rather than model grid resolution/spacing.

The hypothesis for this study was as follows: can WRF, configured with the tropical-suite schemes and initialized with GFS forcing data, adequately reproduce the extreme rainfall event that occurred in Mahalapye District? The findings from the study support this hypothesis: the WRF model was able to capture the broad-scale features and synoptic drivers, including the surface low-pressure system and mid-level trough that induced instability, which agrees with SAWS’s and GFS’s observed weather charts. From this study, we can strongly conclude that the model’s underestimation of rainfall totals points to limitations in cloud microphysics and cumulus parameterization. Studies by [20,22] highlighted that incorporating high-resolution microphysics and data assimilation can reduce these biases from the WRF model.

4. Conclusions

This study assessed the capability of the WRF model in simulating an extreme rainfall event that occurred at the Mahalapye District on 26 December 2023. The results show that the WRF model run at 9 km and a 3 km nested domain is able to reproduce the spatial and temporal distribution of rainfall, but with significant biases. The observed maximum rainfall recorded was 114 mm at the Mahalapye Meteorological station, yet the WRF model simulations significantly underestimated its total rainfall. The findings of the study are consistent with earlier work from other authors, including [22,47], thus indicating that WRF frequently underestimates extreme rainfall magnitudes. However, the model was able to capture the spatial distribution of rainfall reasonably well.

Besides bias, both domains achieved a good probability of detection (POD = 1.00), indicating that the WRF model effectively identified the occurrence of rainfall. A high POD value indicates that the models’ ability to capture rainfall timing and location is impressive, which is especially important in applications of early warning [51]. Considering WRF model bias, the negative bias values (−56.1% to −48.2% for WRF at 3 km and 9 km, respectively) highlight the models’ systematic underestimation of rainfall, which was also observed over stations in Lagos by [47] and in southern Africa by [19]. These findings strongly suggest that the models’ tropical-suite physics configuration captures convective activities during their initiation but struggles to sustain rainfall intensities.

The implications of this study’s findings extend to flood risk management and disaster preparedness in Botswana. Flooding episodes in the country of Botswana are becoming more frequent, with increasing socio-economic impacts [3,4]. This could be due to the impacts of climate change and variability [36]. This study confirms that dynamical downscaling with the WRF model provides valuable predictive information on extreme rainfall but highlights the necessity of model customization for local conditions. The WRF models’ systematic biases identified in this study reinforce the importance of ensemble-based sensitivity testing to provide probabilistic guidance, as suggested in African regional studies [11,47].

In conclusion, future research should focus on multi-physics ensemble simulations, data assimilation of local BDMS observations, and coupling WRF outputs with hydrological models to directly forecast flash floods’ magnitudes and their associated impacts. Furthermore, extending verification beyond a single event to multiple case studies would also provide a more robust assessment of the WRF model reliability across varying synoptic conditions.