Abstract

Despite its significant impacts on climate, environment, and public health, air pollution monitoring in Africa is still sparse. This study developed a machine learning framework of four models—namely SARIMAX, Random Forest, XGBoost, and LightGBM—that were trained on a daily basis using 25 meteorological variables from ERA5 reanalysis to reproduce daily PM₁₀ surface concentrations in Lamto, Côte d’Ivoire. A total of 2225 daily recorded concentrations from 2017 to 2023 were used, with 80% (1780 days) allocated to training the models and 20% (445 days) for prediction. An assessment of ERA5 reanalysis data against in-situ measurements highlighted its ability to reproduce the pattern of meteorological variables such as temperature and relative humidity, despite some biases. Results show that all models were able to reproduce the observed PM₁₀ variations, although they slightly overestimated concentrations during the main wet season (March-April-May-June-July) and underestimated high pollution events in the main dry season (November-December-January-February). Finally, performance analysis revealed that the Random Forest model outperformed, with an R² of 0.78 and RMSE of 23 µg/m³, outperforming the other models. This framework successfully demonstrates the utility of machine learning for air quality prediction in West Africa, with potential for future improvements through bias correction and model combination.

Keywords

Lamto, Prediction, PM₁₀, Machine Learning, Air Pollution, Performance

Share and Cite:

1. Introduction

Human beings, through their activities, have strongly impacted and shaped their environment, including the Earth’s atmospheric composition. Since the Industrial Revolution in the 18^th century, numerous compounds have been released into the atmosphere. These atmospheric compounds are responsible for current global warming, which induces climate change [1] [2]. They contain aerosols, defined as particulate matter in a solid or liquid state suspended in the atmosphere. These aerosols are reported to play a significant role in the climate through their direct, indirect, and semi-direct effects [3] [4]. Their size distribution in the total suspended particles is often classified into three (3) modes: coarse (PM₁₀), fine (PM_2.5), and ultra-fine (PM₁) particles. Their lifetime in the atmosphere is strongly dependent on their mode [5] and generally ranges from a few hours to several weeks. The coarse mode particles tend to be deposited close to their sources, while fine and ultra-fine particles, which contribute to particulate matter (PM), have a longer lifetime and can undergo long-range transport. These particles contribute to air quality deterioration and have numerous impacts on climate, ecosystems, and human health [6]-[8].

Air quality and its impacts in West Africa have drawn attention in several works, revealing increasing air pollution affecting population health [9] [10]. In urban areas, they are often investigated using ground-based observations [11]-[15], satellite [16], and modeling [17] [18]. During the Dynamics-Aerosol-Chemistry-Cloud Interactions in West Africa (DACCIWA) project, numerous studies have highlighted that anthropogenic pollutants emitted from the coastal region of the Guinean Coast were transported to remote areas [19] [20]. As urban areas’ air quality received considerable attention, only a few studies have focused on air quality in rural areas. Investigating daily Aerosol Optical Depth (AOD) in 11 stations across various biomes of Southwest Africa, Niamien [21] found that their AOD ranged from 0.35 in the Sahel area to 0.49 in the Guinea Coast area, and a maximum of 0.53 was observed in the savanna area. Moreover, the International Network to study Deposition and Atmospheric Chemistry in Africa (INDAAF, https://indaaf.obs-mip.fr), has allowed researchers to focus attention on the atmospheric chemical composition of various biomes and rural areas in Africa [22]-[25]. For example, Lamto atmospheric composition was reported to be influenced by emissions from both local and regional sources, particularly from mesoscale sources with a strong anthropogenic signature [26].

Another important aspect of air quality degradation is its strong impact on the environment and human health. According to the World Health Organization (WHO), air pollution is responsible for more than 7 million premature deaths every year, particularly affecting urban and peri-urban areas in developing countries [27]. In Africa, a report on the current situation of air quality highlighted an increase in air pollution across the continent, which was responsible for 250,000 premature deaths in 2018 [28]. Among these air pollutants, aerosols, particularly PM₁₀, are recognized as some of the most harmful components for both human health and ecosystems. For instance, PM₁₀ is linked to respiratory and cardiovascular diseases and contributes to climate imbalance [29]. Moreover, prolonged exposure to high concentrations of PM₁₀ has been associated with increased hospital admissions for respiratory conditions such as asthma, obstructive lung disease, and pulmonary infections. A connection between peaks in PM₁₀ levels and deaths induced by cardiovascular disease has been indicated through strong correlations, highlighting an increase in myocardial infarction, stroke, and hypertension [30]. Despite their hazardous nature, monitoring of these particles remains limited, especially in West Africa, where the high cost of referenced instruments and the scarce infrastructure obstruct continuous monitoring. This lack of data compromises the implementation of effective public policies for air quality management. However, certain scientific initiatives, such as the INDAAF program, have enabled the regular and long-term collection of atmospheric chemistry data from multiple sites in Africa [22]-[25]. Over Lamto, an INDAAF supersite located in the center of Côte d’Ivoire, the atmospheric chemistry composition is monitored. The composition of the atmosphere at this station is the result of complex interactions between natural and anthropogenic sources in a rural area [26]. This makes Lamto an ideal site for studying how particulate compounds interact with meteorological conditions. However, spatiotemporal coverage remains too limited for full operational monitoring of this area. In response to these structural limitations, the integration of new technological approaches is emerging as a promising alternative. Among them, Artificial Intelligence (AI), and more specifically Machine Learning (ML) methods, offer powerful tools to compensate for the lack of data and improve predictive capabilities. The ML algorithms usually can analyze data from various sources, including air quality sensors, satellite imaging, and weather data, to provide real-time information on air quality [31]-[33]. Information about air pollution sources, pollution levels, and attenuation actions can be provided by these algorithms under the conditions they have been trained to recognize patterns and trends in the data [32]. However, all these algorithms do not perform similarly, and their choice may be highly dependent on the nature of pollutants and the purpose of the study. For example, in their health-related research, Dobrea [34] compared other AI-oriented techniques/models for air pollutants and suggested Support Vector Regression (SVR) and AutoRegressive Integrated Moving Average (ARIMA) as the best-performing techniques for time series analysis of PM₁₀ and PM_2.5. Moreover, some works emphasized that hybrid models have better performance and present several advantages for environmental monitoring policy and decision-making [35] [36]. Therefore, by exploiting existing data, these techniques enable the anticipation of changes in pollutant levels such as PM₁₀, thereby supporting environmental management efforts, even in low-resource settings. Their capacity to identify complex, non-linear relationships makes them particularly suitable for environments like Lamto, where multiple factors influence atmospheric processes. In this context, the present study aims to explore for the first time the potential of Artificial Intelligence (AI) models to reproduce and predict PM₁₀ concentrations at Lamto over seven years (2017-2023). To achieve this, four AI models were employed, and their performances were assessed.

2. Materials and Methods

2.1. Sampling Site

Lamto is a station located at 6˚13'N and 5˚2'W, approximately 120 km from Abidjan (Côte d’Ivoire), situated within the humid savanna (Figure 1). The Lamto reserve was established in 1962, covering about 2,617 ha, and has been the subject of numerous research studies [37]-[42]. The reserve contains various measurement observation sites for seismology, infrasound, climatology, Greenhouse Gas (CarboAfrica European project, http://www.carboafrica.eu/), atmospheric mercury, atmospheric chemistry (INDAAF), and ecology. The reserve is often burned around mid-January to remove epigenetic phytomass and conduct ecological experiments [41] [42]. The climate of the Lamto reserve is governed by the West African Monsoon (WAM) and the seasonal shift of the Intertropical Convergence Zone (ITCZ). The annual mean rainfall ranges between approximately 1240 and 1590 mm [40], with four main seasons: a dry season (November to February), a wet season (March to July), a slightly dry season in August, due to monsoon jump, and a slightly wet season from September to November (monsoon retreat). Temperatures are relatively hot throughout the year, averaging around 27˚C, while

Figure 1. (a) Localization map of the Lamto-(Côte d’Ivoire) reserve (6˚31'N-5˚02'W). The green point indicates the Lamto location; numbers in red are the mean annual precipitation (1962-1997) of some synoptic stations, and from north to south, are located the types of vegetation [43]; and (b) zoom on the vegetation of Lamto [44].

relative humidity varies between 65% (minimum in January during the main dry season) and 82% (maximum in June during the main and first rainy season).

2.2. Data

2.2.1. PM₁₀

The PM₁₀ concentration data used in our study were retrieved using the TEOM (Tapered Element Oscillating Microbalance) 1400A, covering seven years from January 1, 2017 to December 31, 2023, at the Lamto station [45]. Lamto is one of the super sites of the INDAAF program (https://indaaf.obs-mip.fr/). INDAAF is a National Observation Service (SNO) of the Institut National des Sciences de l’Univers (INSU) of the Centre National de Recherche Scientifique (CNRS) and is supported by Institut de Recherche pour le Développement (IRD). These PM₁₀ concentration data are available on the INDAAF program website at an hourly scale in micrograms per cubic meter (μg/m³). To reduce the number of missing values, we aggregated the PM₁₀ data to compute daily averages. This approach led to 16.43% missing values across the entire dataset.

2.2.2. Meteorological and Physical Data

Hourly weather and climate parameter data were extracted from the ERA5 database [46] at the grid point closest to the Lamto station. ERA5 is the fifth generation of reanalyses from the European Centre for Medium-Range Weather Forecasts (ECMWF). The data, available from 1940 to the present day, relate to climate and weather conditions on a global scale over the last eight decades. For the present work, 25 meteorological variables (Table 1) have been selected from January 1, 2017 to December 31, 2023. These variables were chosen because they are known to represent weather conditions and atmospheric dynamics. Indeed, these variables relate to wind, temperature, humidity, the planetary boundary layer, and precipitation, which are known to influence the dispersion, accumulation, and leaching of aerosols [47]-[50].

Table 1. List of ERA5 meteorological variables used for the machine learning model training and prediction of PM₁₀ concentration.

2.3. Models, Training, and Evaluation

2.3.1. Models

To achieve the goals of this study, four machine learning models were used (Table 2). Their performances are based on their ability to process linear and non-linear relationships between variables, regression and classification, boosting skills, memory efficiency, etc.

Table 2. Brief description of the different models used in this study.

2.3.2. Model Training and cross-Validation Building Sets

For this study, the data were split into two distinct parts. The first part, representing 80% of the total data length, constitutes a training set. This training set contains 80% of the oldest observations, i.e., starting from January 1, 2017. The second part of the data, representing the remaining 20% of the total data length, lies in the most recent period, i.e., closest to December 31, 2023. This split follows a strict temporal logic in which no random permutation of the data (shuffle) was carried out, which is essential in time series to preserve the coherence of the chronology of events. Respecting the order of the dates avoids any leakage of information from the future to the past at the learning stage, which would distort the model’s actual performance under prediction conditions. The choice of the 80/20 ratio is widely recognised as a good compromise between the quantity of data for training and the robustness of the evaluation [52].

• Model training and cross-validation

Model training is a decisive stage in the modelling chain. The model training objective is to adjust the internal parameters of the algorithms so that they learn, from the training data, to predict the PM₁₀ concentration as a function of the meteorological variables selected. The rigor of this stage directly conditions the quality and robustness of future predictions, as well as the reliability of the assessment. In keeping with the temporal nature of the data, the models were trained solely on the training set (representing 80% of the data), with the remaining 20% being reserved for the test. No random resampling (shuffle) was carried out, in accordance with good practice in time series modelling [53], in order to preserve the chronological order essential for model consistency.

When modelling time series, it is essential to assess the robustness of the models while respecting the temporal structure of the data. Unlike the classic cross-validation method, which assumes the independence of observations and uses random partitions, time series require a specific approach to avoid the biases linked to time dependency. To meet this requirement, the cross-validation method adapted to time series was used, as described by scikit-learn in its official documentation. This method, called TimeSeriesSplit, divides the data into several segments while preserving the chronological order. Each successive segment uses a larger portion of the training data, followed by a test set that is immediately later in time. The use of TimeSeriesSplit provides a more realistic assessment of model performance on unseen data, while avoiding the data leakage problems that could arise with inappropriate cross-validation methods for time series.

The exploratory analysis shows that the year 2023 has 58.63% missing data. Aware of this irregularity, we chose to restrict the study period to the time segment running from 14 January 2017 to 16 February 2023, a period during which data coverage is much more satisfactory. This delimitation made it possible to significantly reduce the overall rate of missing values to 8.99%, a proportion generally considered acceptable for modern imputation methods [54].

• Statistical evaluation

Performance evaluation is a crucial stage in the predictive modelling process, as it enables us to judge the accuracy, robustness, and practical usefulness of the model. Three metrics commonly adopted in time series prediction work were used, namely, the mean absolute error (MAE), the root mean square error (RMSE), and the coefficient of determination (R²).

MAE (Equation (1)) measures the average absolute deviation between the predicted and observed values. It is simple to interpret and not very sensitive to outliers, but it does not consider the dispersion of errors.

$MAE=\frac{1}{n}{\displaystyle \sum }_{i=1}^n\left|y_i-{\widehat{y}}_i\right|$ (1)

where:

• $y_i$ : actual value at time i.

• ${\widehat{y}}_i$ : predicted value at time i.

• $n$ : total number of observations.

RMSE is the square root of the average of the squared deviations between the actual and predicted values. It is shown in Equation (2):

$RMSE=\sqrt{\frac{1}{n}{\displaystyle \sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}$ (2)

This metric gives greater weight to large errors. It is, therefore, particularly useful for detecting models that fail to correctly predict peaks in pollution, a critical aspect of air quality.

Coefficient of determination (R²)

The coefficient of determination, R² given in Equation (3), measures the proportion of the variance in the observed data that the model explains.

$R^2=1-\frac{{{\displaystyle \sum }}_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}{{{\displaystyle \sum }}_{i=1}^n{\left(y_i-\overline{y}\right)}^2}$ (3)

where $\overline{y}$ is the mean of the actual values. An R² value close to 1 indicates that the model explains the variability of PM₁₀ concentrations well, while a negative value means that the model performs worse than a simple average.

3. Results and Discussion

3.1. Assessment of ERA5 Data over Lamto

Time series of daily air temperature at 2 m (Figure 2(a)), daily minimum (Figure 2(b)) and maximum (Figure 2(c)) temperature, as well as relative humidity at the synoptic hours 06:00 UTC (Figure 2(d)), 12:00 UTC (Figure 2(e)), and 18:00 UTC (Figure 2(f)) from 14 January 2017 to 16 February 2023 for the ERA5 and in-situ are compared. Results show that the diurnal cycle of these selected variables is quite well reproduced in the ERA5 reanalysis. However, ERA5 is underestimating the 2 m daily air temperature as well as the daily maximum temperature. The air temperature at 2 m fluctuates between 21.3˚C and 27.9˚C for ERA5 and between 22.7˚C and 33.8˚C for the in-situ. Mean daily air temperature is about 24.5˚C and 28.7˚C for ERA5 and in-situ measurements. Daily maximum temperature varies between 21.5˚C and 27.8˚C with a mean of 24.5˚C for ERA5, and 22.0˚C and 41.5˚C with a mean of 34.2˚C for in-situ measurements, respectively. When considering the daily minimum temperature, ERA5 is slightly warmer than the in-situ measured data at Lamto. ERA5 exhibits maximum temperature ranging from 21.2˚C to 27.5˚C with a mean of 24.3˚C, while the in-situ values range from 16 to 34.9˚C. It can be noticed that the extreme events characterized by a sharp increase or decrease in the temperature magnitude are poorly reproduced in the ERA5 data. This indicates a poor performance of using ERA5 to investigate extreme events over the Lamto station.

Figure 2. Timeseries comparison of daily air temperature at 2 m (a), daily minimum (b), and maximum (c) temperature, and relative humidity at the synoptic hours 06:00 UTC (d), 12:00 UTC (e), and 18:00 UTC (f) for the ERA5 and in-situ data between 14^th January 2017 and 16^th February 2023.

Analyzing the relative humidity comparison between ERA5 and in-situ recorded at synoptic hours, it can be observed that the main features (diurnal cycle, seasonality, etc.) match and the magnitude between the two datasets is in the same range. At 06:00 UTC, it varies from 51% to 100% with a mean of 95% for ERA5 and from 43.6% to 100% with a mean of 95% for the in-situ data. At 12:00 UTC, relative humidity fluctuates from 19.2% to 97% with a mean of 65% for ERA5, and for the in-situ, values range from 15.3% to 100% with a mean value of 66.7%. The mean relative humidity from the ERA5 reanalysis at 18:00 UTC is about 73.6% (range of 13.5% to 99.9%), and for the in-situ, the mean humidity is about 69.7% (range of 16.9% to 99.1%).

Figure 3 shows the Spearman correlation between the in-situ and ERA5 variables. Quite good and strong correlations (r varying from 0.48 to 0.73) are found between the temperature variables, and quite moderate correlations (r varying from 0.22 to 0.65) are found between the relative humidity variables, significant with p-values less than 0.01. These correlations between ERA5 and in-situ measured variables were like those reported by Bodjrènou [55] when assessing ERA5 hydrological variables with in-situ measurements over Benin. There is also moderate to strong anti-correlation between temperature and relative humidity, indicating that the relative humidity at Lamto station tends to decrease when temperature increases, as warmer air has a greater capacity to hold water vapour than cooler air [56].

Figure 3. Heat map of the Spearman correlation between the in-situ and ERA5 of daily mean air temperature at 2 m, daily minimum and maximum temperature, relative humidity at the synoptic hours 06:00 UTC, 12:00 UTC, and 18:00 UTC between 14^th January 2017 and 16^th February 2023.

Overall, the ERA5 data capture the diurnal and seasonal cycles of climate and meteorological features of Lamto. As an evaluation of ERA5 data over the Lamto station using in-situ observed data did not exist in the literature, it appeared important for us to perform this quick evaluation on the ERA5 data, as it offers the advantage of a wider range of meteorological variables needed for the use of Artificial Intelligence (AI) tools in this work.

3.2. Exploratory Analysis

Figure 4 shows the time series evolution of daily PM₁₀ concentrations at the Lamto station, marked by a strong seasonal cycle, with a peak occurring during the dry season (November to March). Minimum PM₁₀ concentrations are recorded during the rainy season (April to October). In contrast to a bimodal rainfall pattern observed at Lamto, the PM₁₀ concentration pattern is characterized by a single peak occurring during the dry season. Moreover, the highest PM₁₀ concentration peaks were recorded in 2018, with values reaching 888.9 μg/m³ on 1^st January. The year 2021 recorded relatively low PM₁₀ concentrations during the dry season compared to the other years. The dry season in Lamto is often impacted by pollutants from both local and long-range sources. During the dry season, Lamto station and surrounding areas are subject to biomass burning fires [57], which are set by different groups of people for different reasons (research purposes, maintaining savanna landscapes, grazing and agriculture, hunting, ...). Moreover, N’Datchoh [58] showed that biomass burning is prevalent over the West African region from November to December, while analyzing satellite fire products from SPOT VEGETATION. In addition to these local and long-range sources from biomass burning activities, this region of West Africa is also reported to be impacted by dust aerosols transported from the Sahel and Saharan region to the Guinean Gulf by the Harmattan winds. During the main rainy season, minimum PM₁₀ concentrations are recorded with daily concentration values well below 50 μg/m³. Furthermore, these PM₁₀ daily concentration values maintain their lowest values during the little dry season. This little dry season in the region along the Guinean coast is often due to the WAM jump, which allows the WAM rain belt to be located over the Sahel region [59].

Figure 4. Daily PM₁₀ concentration time series, with different colours corresponding to a specific year.

Table 3 summarizes the descriptive statistics for the raw PM₁₀ database at Lamto. A total of 2225 days were analysed. An average of daily PM₁₀ concentrations of about 42.79 μg/m³ is associated with a very high standard deviation of about 59.5 μg/m³, as well as a median of 22.15 μg/m³, indicating a high variability and lack of balance in PM₁₀ concentrations in Lamto. This type of distribution, asymmetrical and spread out towards the high values, reflects the presence of occasional high pollution events.

Table 3. Daily PM₁₀ concentration at Lamto dataset basic statistics overview between 2017 and 2023.

3.3. Decomposition of the PM₁₀ Time Series

Figure 4 presents the original PM₁₀ concentration time series (Figure 5(a)) and its decomposition into trend (Figure 5(b)), seasonality (Figure 5(c)), noise (Figure 5(d)), and their transformation into a logarithmic function (Figures 5(e)-(h)). The logarithmic function was applied to the daily PM₁₀ concentration for the purpose of smoothing and attenuating the existing high variability observed in the data, making variation easier to analyse, as suggested by Hyndman and Athanasopoulos [60]. The overall trend (Figure 5(b) and Figure 5(f)) exhibits phases of high and low daily PM₁₀ concentrations, which are the result of complex interactions between environmental, climatic, and meteorological factors as well as human activities. A comparison between the original (Figure 5(b)) and logarithmic function (Figure 5(f)) shows a smoothing, allowing detection of changes that are crucial in the use of linear models. The seasonal decomposition (Figure 5(c) and Figure 5(g)) shows the seasonal cycle in the daily PM₁₀ concentrations, which are dominated by higher values of daily PM₁₀ concentration during the dry season compared to the wet season. Here, the use of the logarithmic transformation allows smoothing of the seasonal cycle while capturing the annual and seasonal cycles. Residual or noise components (Figure 5(d) and Figure 5(h)) characterize all daily PM₁₀ concentration data, which cannot be explained either by trend or seasonality. Since SARIMA needs stability to provide accurate results in terms of its output, the logarithmic transformation allows smoothing of the data and facilitates its machine learning process, which can be affected by very high values such as extreme events [61]. The presence of structured trends and well-defined seasonality favored the use of classical models such as SARIMAX, which can take explicit account of such trend and seasonality. However, the unpredictable nature of certain variations and the possible non-linear interactions between daily PM₁₀ concentration and meteorological variables favoured the consideration of more flexible machine learning models, such as XGBoost, Random Forest, or LightGBM. These latter models are known for their robustness in dealing with noisy and highly variable data and their ability to capture complex relationships [62] between variables.

Figure 5. Time series of PM₁₀ concentrations and logarithmic transformations (log): Original ((a) and (e)), Trend ((b) and (f)), Seasonality ((c) and (g)), and Noise ((d) and (h)).

A Shapiro-Wilk test recommended for moderate to large samples was conducted in order to check whether the daily PM₁₀ follows a normal distribution. The Shapiro-Wilk test results reveal that all the variables considered here deviated significantly from normality, with a very low statistic for PM₁₀ (W = 0.51) and a p-value close to zero (≈1.12 × 10⁻⁵⁹) (W = 0.51) and a p-value close to zero (≈1.12 × 10, implying that classic parametric tests were less appropriate for this work. Therefore, a non-parametric Kruskal-Wallis test was conducted to check monthly and yearly variability significance in the daily PM₁₀ dataset. This allows confirmation of the existence of strong monthly variability and seasonality of all variables considered. An ADF (Augmented Dickey-Fuller) stationarity test, which is a must for any time series study [63], was conducted. These test results indicate that the time series are stationary, meaning that their statistical properties (mean, variance) remain constant over time.

Figure 6 presents the seasonal pattern for PM10 concentration climatology between 2017 and 2023.

Figure 6. Boxplots showing the seasonal pattern of PM₁₀ concentrations from 2017 to 2023.

3.4. Explanatory Variables

The final selection of explanatory variables and their optimal lags is presented in Table 4. Time lags of 1 to 3 days [PM₁₀(t-1), PM₁₀(t-2), PM₁₀(t-3)] were included as additional explanatory variables. This decision was motivated by two preliminary analyses. (i) An examination of the autocorrelation (ACF) and partial autocorrelation (PACF) functions revealed significant dependence up to a lag of three days. (ii) A comparative performance evaluation showed that adding lags beyond 3 days lag (d-3) did not significantly improve RMSE and R² metrics, while increasing the risk of overfitting. This choice is also consistent with several previous work [64] that have demonstrated the significant contribution of delayed PM₁₀ values in predicting future concentrations. To avoid any leakage of information from the future to the past, which would distort the predictions, a breakdown of the time series into two sets was performed: the training data (Train) covers the period from 14 January 2017 to 28 November 2021, i.e., 1780 days, and the test data runs from 29 November 2021 to 16 February 2023, representing 445 days. This breakdown respects the natural temporal logic of the series, a fundamental criterion in the context of forecasting. This breakdown can be seen in Figure 7. Figure 7(a) shows the separation of the daily PM10 concentration series into train (blue color) and test before filling the missing values. Figure 7(b) shows the daily PM₁₀ concentration with the missing values filled using the median of similar days in previous and following years to consider the strong seasonality observed in the dataset. For example, the missing values on 10 July 2019 are interpolated using the median for 10 July in 2017, 2018, 2020, etc. This method inspired by the work of Weed [65] makes it possible to retain seasonal effects without introducing time distortion. Although this method is not the most common in the literature, it is based on a simple and solid logic, which takes advantage of seasonal repeatability from one year to the next. It is particularly well suited to environmental data, where seasonal effects are often very marked.

Table 4. Explanatory variables and optimal time lags used in this work.

Figure 7. Daily PM₁₀ surface concentration breakdown into train from 14^th January 2017 to 28^th November 2021 (1780 days) and test runs from 29^th November 2021 to 16^th February 2023 (445 days) for (a) original data and (b) missing data fill-up.

To handle missing values in daily PM₁₀ concentrations, we used the median of observations from the same calendar day in other years of the study. This method was selected for its proven ability to preserve the temporal characteristics of time series. However this way of handling the missing data may potentially bias the model training by reducing the natural variance of the data and smoothing out actual pollution peaks. This may lead the model to underestimate the amplitude of surface PM₁₀ concentration peaks during high pollution events. As a result, test errors could be underestimated during validation, as the model would be evaluated on smoothed data that does not fully reflect the diversity and intensity of the observed conditions, which could limit its ability to generalize to new data.

3.5. Separation of Time Series and Handling of Missing Values

To assess the reliability and ability of the Machine Learning models to generalize well the predictions of PM₁₀ concentrations, a cross-validation was done. The model was trained on past data to make predictions about a future period, without ever using information from the future. Three models were used, namely Random Forest, XGBoost, and LightGBM. For each model, performance was measured using the root mean square error (RMSE). Mean errors obtained are summarized in Table 3, highlighting that the Random Forest model performs better, with an average RMSE of 15.94. It is closely followed by LightGBM (16.44), then XGBoost (17.03). These scores indicate that Random Forest seems to better capture the dynamics of the daily PM₁₀ concentration, while remaining stable across the different periods tested. However, the SARIMAX model was not included in this cross-validation. The SARIMAX model relies heavily on the temporal dependence between observations, and therefore was evaluated separately, using a simple chronological breakdown between training and testing.

3.6. Final Evaluation and Model Performance

Figure 8 shows the original (blue) and predicted time series modelled (orange) by daily PM₁₀ surface concentrations from January 2022 to March 2023 for SARIMAX (Figure 8(a)), Random Forest (Figure 8(b)), XGBoost (Figure 8(c)), and LightGBM (Figure 8(d)). The hyperparameters used for each model are presented in the supplementary material (Table S1). It should be noted, that no performance optimisation tests were carried out on in this present work. We used default configurations for each of the models. Overall, all four models were able to capture the general trend of the PM₁₀ surface concentration in Lamto. However, a close analysis of Figure 8 reveals some slight differences between the models themselves and between the models and the in-situ observations. For example, the PM₁₀ concentrations range from 10.19 (minimum) to 210.01 (maximum) µg/m³ for the SARIMAX model, 10.48 (minimum) to 171.88 (maximum) µg/m³ for the Random Forest model, 12.47 (minimum) to 190.37 (maximum) µg/m³ for the XGBoost model, and 11.68 (minimum) to 171.20 (maximum) µg/m³ for the LightGBM model, while the observations are within 8.00 (minimum) and 266.57 (maximum) µg/m³. All the models tend to overestimate the PM₁₀ minimum concentration, while they underestimate the maximum compared to the observations. Moreover, all models tend to slightly overestimate PM₁₀ concentrations compared to observations during the wet season. Furthermore, in terms of capturing the high variability in PM₁₀ concentration peaks associated with pollution events, some slight lags seem to be produced in all the models except the Random Forest, which seems

Figure 8. Original (blue) and predicted (orange) time series simulated by the (a) SARIMAX, (b) Random Forest, (c) XGBoost, and (d) LightGBM models of daily PM₁₀ surface concentrations from January 2022 to March 2023.

Table 5. Statistical indicators RMSE, MAE, and R² provided at annual and seasonal time scales for each of the four models used in this study.

to follow more closely. Therefore, it appears necessary to evaluate the performances of the different models that are measured using indicators such as RMSE, MAE, and R². These indicators are summarized in Table 5 for the entire predicted timeseries but also per season. From Table 5, it can be seen that the SARIMAX model is the least performant model with a limited capacity to capture changes in the daily PM₁₀ concentrations at Lamto, as it scored lower performance, with a RMSE of 28.52 μg/m³, a MAE of 16.52 μg/m³, and a R² of 0.64. These relatively poor scores underline that the present models (Random Forest, XGBoost, LightGBM, and SARIMAX) have poor capacity in reproducing daily PM₁₀ surface concentration, which may be explained by its high variance and the complex interaction between meteorology, local, and long-range factors that influence the PM₁₀ loaded in the Lamto atmosphere. PM₁₀ concentrations in Lamto are influenced by complex non-linear phenomena as well as irregular exogenous factors, which are difficult to model within a purely statistical framework.

Conversely, the Machine Learning models show a potential possibility to be adapted and to integrate complex atmospheric processes involving local and long-range factors under the condition of providing good and robust observations, which may be used to train and test them. Therefore, the Random Forest model scored a better overall performance with an RMSE of 23.12 μg/m³, an MAE of 12.35 μg/m³, and an R² of 0.78. Furthermore, an RMSE of 23.12 indicates an average prediction error of 23.12 μg/m³ compared with the actual values, while an MAE of 12.35 underlines that the average absolute difference between the predicted and observed values is 12.35 μg/m³. The coefficient of determination R² = 0.78 indicates that 78% of the variability in the daily PM₁₀ surface concentrations may be explained by the model, which is satisfactory but also suggests margins of error. Therefore, in general, the pattern of the daily PM₁₀ concentration is better reproduced by the random forest, despite some bias. For example, the predicted PM₁₀ values during the dry season (November to March) are less important than those observed. These biases are associated with specific events of PM₁₀ pollution, which are more frequent during the dry season. The relatively low daily PM₁₀ surface concentration during the wet season is quite well reproduced in both pattern and magnitude, highlighting the ability of machine learning to better assimilate periods characterized by weak changes.

Variables permutation importance for the Random Forest, LightGBM, and XGBoost are shown in Figure 9. This post-modeling analysis confirms the relevance of preliminary selection of variables based on mutual information and cross-correlation. The results of that PM10_lag1 predominance highlights a temporal self-dependence as the predominant mechanism in predicting PM₁₀ surface concentration in Lamto. Also, the PM10_lag2 and PM10_lag3 delays confirm this multi-day persistence, may be related to the dry season where rainfall is less frequent, less washout of atmosphere occurring and more pollutant particularly PM₁₀ holds into atmosphere. When considering, meteorological variables, boundary layer height (var8_blh_lag1), humidity (var4_d2m_C_lag1), temperature (var5_mx2t_lag1), and latent heat flux (var14_slhf_lag1) emerge as consistent secondary modulators, influencing vertical dispersion, hygroscopic particle growth, convection, and turbulence, respectively. Finally, the Month_lag1 variable robustly captures the seasonality documented in Lamto. This multi-model analysis validates both the robustness of the predictors and the complementarity of our variable selection approaches. Finally, because of the SARIMAX model’s structure, variable permutation is not possible; therefore, permutation importance could not be computed and was excluded from this analysis.

Figure 9. Variables permutation importance for (a) Random Forest; (b) XGBoost; and (c) LghtGBM.

3.7. Discussion

This work focused on the use of Machine Learning (ML) tools for predicting PM₁₀ concentrations. To achieve this goal, the methodology applied consisted of four ML models, namely SARIMAX, Random Forest, XGBoost, and LightGBM models, to train them over 80% of the total filtered PM₁₀ data length. The remaining 20% of the observed PM₁₀ data was used for testing the model outputs. The Random Forest was identified as the best-performing model among these four models, with estimated biases of 23.12 µg/m³. These observed biases between original and predicted daily PM₁₀ surface concentration at Lamto may be explained, on the one hand, by the complex mixture of sources influencing its chemical composition [22] [26] [40]. Lamto chemical composition is the result of a combination of local (biomass burning, domestic fires, biogenic emission, ...) and remote (long-range transport of dust from the Northern Hemisphere of Africa during the dry season, biomass burning, sea salt, ...). The dry season months are characterized by high variability in pollution events, as all local and long-range transport of Sahel and Saharan dust transportation toward the Guinean Gulf [66] are also prevalent. On the other hand, biases observed between ERA5 and the in-situ measurement may have contributed to all models underestimating pollutant events maxima. A combination of in-situ measurement (when these are available) and ERA5 data for the machine learning models may contribute to reducing these observed maxima underestimations observed in all models used. Thus, this low performance of the model during the dry season may be related to the high variability in daily PM₁₀, which contributes directly to the increase in model error and cold biases observed with the ERA5 meteorological temperature variables. During the wet season, the washout of the atmosphere by rain contributes to fewer high air pollution events, thus PM₁₀ loads into the atmosphere. This led to the higher scores observed by all models, especially the Random Forest. This finding highlights the current difficulties and limitations of applying logical tools such as mathematics through modelling to complex environmental processes involving meteorology and climate, which are highly impacted by human socio-economic activities such as air pollution. Arowosegbe [67] used Random Forest models to impute missing daily PM₁₀ concentrations in South Africa between 2010 and 2017 using spatio-temporal predictors including meteorological variables, land cover, elevation, and proximity to roads. They found an average R² of 0.78 at the national level, 0.70 at the provincial level, and 0.55 at the level of each site, thus demonstrating the effectiveness of the Random Forest model in an African context despite the scarcity of data. The temporal component of the national models explained up to 78% of the variability in concentrations. Moreover, Adnane [68] in Morocco used Nonlinear Autoregressive Neural Networks with Exogenous Inputs (NARX) models to predict hourly PM₁₀ concentrations in Agadir, at 1 h and 24 h horizons. The best results were obtained with the 24h horizon model, combining meteorological variables and secondary pollutants (CO, SO₂, O₃) as inputs. The model achieved a correlation coefficient of 93.75%, i.e., an approximate R² of 0.88, illustrating the added value of pollution data combined with meteorological variables in improving predictive performance.

As hybrid models have been reported to have better performance for environmental monitoring policy and decision-making [35] [36], such methods need to be evaluated in the limited data context of sub-Saharan Africa. As AI is emerging as a necessary transversal tool in many research areas, the African continent must integrate such a tool in environmental process monitoring and management. Applying AI to their AirQo low-cost sensors in Africa, Bainomugisha [69] highlighted a set of digital solutions for the environmental air pollution challenges from custom-designed low-cost air quality monitors, deployment methodology, AI-powered digital tools, and a framework for citizens’ and leaders’ engagement. Finally, leveraging AI and data science to mitigate the respiratory health impacts in the context of climate change in Africa, Sowunmi [70] emphasized pilot AI-driven health monitoring systems in major cities with clear indicators, such as improved weather forecasts, reductions in respiratory symptoms or hospitalizations, and improved healthcare access.

4. Conclusions

Daily PM₁₀ surface concentrations in Lamto over seven years (2017-2023) were investigated using a machine learning tool. To achieve this, firstly, the ERA5 daily data were assessed with in-situ measurements recorded at the Lamto station. This assessment shows that, despite some biases, ERA5 can reproduce a fairly diurnal cycle of temperature and relative humidity for Lamto. Good and strong correlations (r varying from 0.48 to 0.73) are found between the ERA5 and in-situ measurements of temperature variables. Also, moderate correlations (0.22 to 0.65) exist between the relative humidity variables of these two datasets.

Using 25 meteorological variables from daily ERA5 and PM₁₀ surface concentration, we applied machine learning to 80% (14^th January 2017 to 28^th November 2021, i.e., 1780 days) of the daily PM₁₀ surface concentration to train four machine learning models (SARIMAX, Random Forest model, XGBoost, and LightGBM), while 20% (from 1^st January 2022 to March 2023, i.e., 445 days) were used to validate their predictions. All the models reasonably reproduced trends and patterns of PM₁₀ daily concentration compared to the in-situ observations, despite some biases. These biases consisted of a slight overestimation of the daily PM₁₀ concentration during the wet season, while underestimating high pollution days during the dry season. The performances of these four machine learning models were evaluated using indicators such as RMSE, MAE, and R². The Random Forest model achieved an R² of 0.78 and an RMSE of 23 μg/m³, surpassing the performance of the SARIMAX model (R² of 0.64 and an RMSE of 28.54 μg/m³), XGBOOST (R² of 0.74 and an RMSE of 24.54 μg/m³), and LightGBM (R² of 0.76 and an RMSE of 23.70 μg/m³). This model’s performance helped identify the Random Forest as the best performing model in reproducing the daily PM₁₀ surface concentration at Lamto. This performance was better during the wet season, when daily PM₁₀ presents less variability with low pollution levels. Despite the high variability associated with pollution events depending on local and remote factors, the pattern was captured despite a lower high pollution magnitude. Moreover, the low performance of the machine learning models during the dry season may be related to this high variability in daily PM₁₀ and biases observed with ERA5 meteorological data, though we did not investigate the error sources and their impacts on the work. Future work will focus on the contribution of bias correction into ERA5 as well as the combination of models to improve the machine learning model performance in capturing environmental processes over Lamto and other places of the country and the continent.

Acknowledgements

The authors would like to acknowledge the INDAAF (International Network to study Deposition and Atmospheric chemistry in Africa) project. INDAAF is supported by the INSU (Institut National des Sciences de l'Univers)/CNRS (Centre National de la Recherche Scientifique), IRD (Institut de Recherche pour le Développement) and the research infrastructure ACTRIS-FR registered on the Roadmap of the French Ministry of Research. We also express our gratitude to all INDAAF local technicians for their maintenance and sampling work. We would also like to thank the JEAI PATI (Jeune Equipe Associée de l’IRD Physico-chimie Atmosphérique et Impacts) project under the sponsorship of the IRD for their generous financial support in the publication of this article.

Data Availability

The INDAAF data can be retrieved from the website https://indaaf.obs-mip.fr/. Interested persons may contact [email protected] or [email protected] for these data.

Author Contributions

Conceptualization and methodology, N.E.T., M.G.O., and C.E.; Analysis experience, N.E.T., M.G.O., C.E., and A.B.; writing—original draft preparation, N.E.T., M.G.O., A.B., C.E., K.K., F.Y., M.D., M.L.D., G.N.S.K., and V.Y.; writing—review and editing, N.E.T., M.G.O., A.B., C.E., K.K., F.Y., M.D., M.L.D., G.N.S.K. All authors have read and agreed to the published version of the manuscript.

Supplement Material

Table S1. Hyperparameter information used for each of the model use in this study. Note that, nosystematic investigation for hyperparameters (using grid or cross-validation) was performed, as we focused in this study on basic comparison of models with their default configuration.

Conflicts of Interest

The authors declare no conflicts of interest regarding the publication of this paper.

References