https://orcid.org/0009-0005-6525-417X
Figures
Abstract
In the field of environmental health, assessing air pollution exposure has historically posed challenges, primarily due to sparse ground observation networks. To overcome this limitation, satellite remote sensing of aerosols provides a valuable tool for monitoring air quality and estimating particulate matter concentration (PM) at the surface. In this study, we employ two predictive models to estimate Aerosol Optical Depth (AOD) levels over Ghana and selected localities from January 2003 to December 2019. Our investigation focuses on evaluating the capabilities of multiple linear regression (MLR) and artificial neural network (ANN) models in predicting AOD levels. Additionally, we introduce a novel approach to constructing the MLR model by leveraging the ANN architecture. These models utilize meteorological variables as input, to facilitate accurate predictions. Despite Ghana’s alarming air pollution health ranking and its substantial role in mortality, routine monitoring remains sparse. This research contributes a comprehensive sixteen-year assessment (2003-2019) of AOD at a 3 km resolution, obtained from MODIS Aqua and Terra satellites. The findings indicate that the southwestern part of the country displays elevated aerosol levels compared to other major cities. Given the region’s dense vegetation, this phenomenon can be attributed to biogenic emissions. Additionally, many small cities within this area are recognized as hotspots for surface mining operations, potentially contributing to increased local dust loadings in the atmosphere. Notably, the MLR model, implemented using the ANN model structure, outperformed the other utilized models. This endeavor aims to unravel the spatiotemporal distribution patterns of aerosols across Ghana, and its major urban hubs.
Author summary
Air pollution from fine particulate matter is a growing public health concern in Ghana, yet ground-based monitoring is limited. In this study, we used satellite data and machine learning techniques to estimate Aerosol Optical Depth (AOD), a key proxy for air pollution, over Ghana from 2003 to 2019. We applied multiple linear regression (MLR) and artificial neural networks (ANN) to predict AOD using meteorological and energy balance variables. To improve performance, we developed a hybrid ANN(MLR) model that structurally mimics linear regression within a neural network framework. This model consistently outperformed traditional MLR and ANN methods, especially in densely populated cities like Accra and Takoradi. Our analysis revealed persistently high aerosol levels in southwestern Ghana, likely driven by biogenic emissions, illegal mining, and industrial activity. Despite challenges such as data gaps and missing observations, our models were able to provide reliable spatial and temporal estimates of aerosol patterns. This work offers a valuable tool for air quality assessment in regions lacking routine monitoring and demonstrates how satellite observations combined with machine learning can bridge critical environmental data gaps in West Africa.
Citation: Gilbert J, Aryee JNA, Adjei MJ, Mensah C, Quagraine KT (2025) Machine learning-based assessment of aerosol optical depth over Ghana, West Africa using MODIS satellite data. PLOS Clim 4(8): e0000651. https://doi.org/10.1371/journal.pclm.0000651
Editor: Muhammad Irfan Ashraf, University of Sargodha, PAKISTAN
Received: January 11, 2025; Accepted: July 29, 2025; Published: August 29, 2025
Copyright: © 2025 Gilbert et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Publicly available datasets were analyzed in this study. The MODIS data can be accessed from: (https://giovanni.gsfc.nasa.gov/giovanni/). The ERA-5 datasets can be accessed from: (https://cds.climate.copernicus.eu/).
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Air quality has emerged as a critical environmental and public health concern globally, largely due to increased atmospheric aerosols with diverse compositions and sizes from different sources. Aerosols are airborne particles suspended in the atmosphere, encompassing various compositions and sizes [1,2]. These particles emerge from two primary sources: direct emission of primary particulate matter (PM) and secondary particle formation from gaseous precursors. Black carbon (BC), primary biological aerosol particles (PBAPs), sea salt spray, and, mineral dust are some examples of primary aerosols. On the contrary, secondary aerosols arise from processes of sulfate, nitrate, and ammonium formation [3]. Aerosols directly influence the atmosphere’s energy balance through radiation scattering and absorption [4]. Furthermore, they act as nuclei for cloud formation [5], and indirectly impact atmospheric heat by absorbing radiation, contributing to reduced low cloud cover [6]. Consequently, aerosols impact the earth’s hydrological cycle [7] and, to a considerable extent, food security [8]. Fine PM with diameters below 2.5 (PM2.5) have been linked to negative health outcomes, such as immediate to chronic effects [9]. These include aggravated respiratory symptoms [10–12], worsened asthma [13,14], heightened cardiovascular diseases [15,16], diminished lung function ([17]), and increased premature mortality linked to heart or lung conditions [18,19].
The surge in vehicle numbers and land use transformations, largely driven by rapid urban growth in Ghana’s major cities [20], has led to escalating PM2.5 concentrations. These concentrations are significantly affected by local traffic emissions, land use practices like bush burning, and industrial discharges. Rapid and precise assessment of the spatiotemporal distribution of PM2.5 [21,22] at a finer resolution can enhance the accuracy of health outcome studies related to PM2.5 [14], especially when conducted on a local spatial scale. However, PM2.5 ground monitoring stations are usually limitedly distributed worldwide, especially for developing countries such as Ghana and its neighbors. This may be a result of resource constraints because establishing and maintaining a network of monitoring stations requires financial investment for equipment, infrastructure, and personnel, which may be challenging for countries with limited budgets. Furthermore, the spatio-temporal variation of PM2.5 is complex, influenced by a combination of factors, including local emissions from various sources, meteorological influences, topography, and seasonal patterns such as the movement of the trade winds over Ghana. In addition to natural sources such as Saharan dust, anthropogenic activities such as illegal timber harvesting also contribute to emissions in Ghana, particularly in forest-rich regions in the south and west [23,24].
Long-range transport from diverse geographical locations and chemical transformations further contribute to this complexity. Without continuous monitoring of PM2.5, our ability to rapidly assess, model, and forecast PM2.5 levels for Ghana, particularly over the major cities, is severely limited. As an alternative, satellite retrievals are used since they have a wider capture and provide information in most instances even for remote and inaccessible locations. The most relevant satellite-derived parameter for assessing PM2.5 concentration levels is Aerosol Optical Depth (AOD) [25–28]. AOD measures the drop in light intensity caused by aerosol scattering and absorption throughout the atmospheric column [29]. This metric directly reflects the extent of aerosol presence, providing valuable insights into the overall concentration of optically active particles within each geographical location [30]. Exposure to elevated aerosol levels has been strongly linked to adverse health outcomes, including respiratory diseases such as asthma and chronic obstructive pulmonary disease (COPD), as well as cardiovascular conditions like arrhythmia and ischemic heart disease [31]. Fine particulate matter (PM2.5), which AOD is often used to proxy, penetrates deep into the alveolar regions of the lungs, triggering inflammation and oxidative stress [10–12]. Epidemiological studies have shown that even small increases in PM2.5 concentrations are associated with heightened risks of hospital admissions and premature mortality. In regions like Ghana, where ground-based monitoring is limited, AOD provides a useful satellite-derived indicator of population-level exposure to air pollution.
In the domain of satellite-based aerosol retrievals, AOD datasets are predominantly sourced from two key satellite observations: the Moderate Resolution Imaging Spectroradiometer (MODIS) [32–34], and the Terra Multi-angle Imaging SpectroRadiometer (MISR) [35,36]. MODIS stands out, utilizing Aqua and Terra satellites. Its widespread adoption is owed to its impressive characteristics: a wide swath width covering 2330 kilometers and near-global coverage every 1 to 2 days [37]. MODIS aerosol retrievals rely on three main algorithms: Dark Target (DT), Deep Blue (DB), and the combined Dark Target and Deep Blue (DTB) algorithms. Further discussion on these algorithms is provided in section “MODIS”. This present study utilized the DTB product, based on the DT and DB algorithm. AOD has consistently demonstrated strong correlations with PM measurements [38–40]. Conventional modeling approaches have predominantly relied on chemical transport models (CTMs) and land-use regression (LUR) models, for modeling PM2.5 levels while incorporating AOD as a key predictor [41,42]. Nevertheless, it is worth noting that Land Use Regression models (LURs) have inherent limitations in capturing temporal fluctuations, while CTMs may exhibit discrepancies when employed in isolation [43]. Various statistical methods, spanning from simple univariate regression to complex non-linear models have been developed for the estimation of PM [29,43–46].
Machine learning (ML) methods have gained popularity in the study of aerosol dynamics and air quality forecasts [43,46–48]. This shift in approach highlights the need for further research to explore the potential of ML in advancing our understanding of aerosol dynamics and improving air quality predictions. The application of advanced algorithms enables non-parametric exploration of the complex relationship between predictor variables and measured pollutant concentrations [49,50]. The current investigation focuses on understanding aerosol dynamics in Ghana, a nation experiencing enhanced population growth and economic expansion [51].
Despite the global rise of machine learning (ML) in air quality modeling, its application to Aerosol Optical Depth (AOD) prediction in Ghana remains limited. This gap is critical given Ghana’s exposure to multiple aerosol sources, including Saharan dust intrusions during the Harmattan season. In the absence of dense ground-based PM2.5 monitoring networks, satellite-derived AOD offers a valuable proxy for assessing spatial patterns of air pollution exposure.
This study addresses this gap by combining spatial diagnostics with ML-based AOD prediction. First, we analyze the spatial and temporal distribution of AOD in Ghana using TERRA and AQUA products, highlighting regions with elevated aerosol burdens and capturing diurnal variation. Second, we assess the performance of machine learning models (MLR (see section “Multiple linear regression (MLR)”), ANN (see section “Multi linear regression (MLR) using ANN”), ANN(MLR) (see section “Artificial neural network (ANN)”)) in predicting AOD from climate and meteorological variables. The models are applied at national and regional scales, with a focus on urban centers identified in the initial spatial analysis. Through this integrated framework, our aim is to contribute to AOD-based air quality assessments and demonstrate the utility of ML approaches in the West African context.
Dataset and methodology
Study area
The study was conducted in Ghana (Fig 1), located on the West African Guinea Coast, a region that experiences a monsoonal climate characterized by distinct dry and wet seasons [52]. These seasonal variations not only drive rainfall patterns but also have significant implications for air quality across the country. The West African Monsoon (WAM) influences the dispersion and concentration of pollutants, as seasonal shifts in wind patterns affect the transport of emissions from both local and transboundary sources [53]. For instance, during the wet season, atmospheric cleansing processes, such as wet deposition and enhanced convection, reduce particulate matter and gas concentrations in the air, improving air quality. Conversely, the dry season is associated with stagnant air masses and reduced rainfall, leading to the accumulation of pollutants, including dust from the Sahara, and emissions from anthropogenic sources like vehicles and industry [54]. The movement of the Inter-Tropical Discontinuity (ITD) also plays a role in modulating air quality [54]. During the dry Harmattan season, the north-easterly winds bring dry, dusty air from the Sahara, which can significantly elevate particulate matter concentrations in the atmosphere, leading to poor air quality.
The top-right panel highlights the Western Region, with a focus on Takoradi, a coastal urban area where high aerosol levels were observed. The bottom panel zooms into the Greater Accra Region, encompassing Accra, another major metropolitan area with elevated AOD values. These regions were selected for both national-scale and regional-scale AOD modeling using MODIS AOD data at approximately 10 km spatial resolution. Base map shapefile from OCHA via humanitarian data exchange (https://data.humdata.org/dataset/cod-ab-gha), licensed under CC BY 4.0.
Ghana is notably one of the fastest-developing countries on the African continent [55]. The nation boasts four key urban centers: Accra, Kumasi, Takoradi, and Tamale. However, this rapid economic growth coupled with population growth, industrialization, and a surge in vehicular density has contributed to incremental growth in air pollution in the country, particularly in these urban centers. Furthermore, findings from [56] indicate that a considerable number of vehicles imported into Ghana are aging, receive limited maintenance, and consequently contribute to elevated emission levels. As of September 2018, over 28,000 deaths in Ghana were linked to air pollution [57,58]. The World Health Organization (WHO) reported that Ghana’s annual average PM2.5 level in 2016 was 31.1 , well above the recommended guideline of 10
[57].
Dataset
MODIS.
MODIS serves as the pivotal instrument aboard NASA’s Terra and Aqua satellites, launched in December 1999 and May 2002, respectively [59]. Terra follows an orbit from north to south across the equator in the morning, whereas Aqua’s orbit goes from south to north over the equator in the afternoon. With Terra MODIS and Aqua MODIS working in tandem, they achieve a comprehensive view of the Earth’s surface approximately every one to two days. This remarkable feat is achieved by capturing data across 36 spectral bands, spanning wavelengths from 0.4 to 14.385 . The visual richness of MODIS imagery emerges with spatial resolutions of 250m, 500m, and 1km [60–63]. Of noteworthy significance are the specific channels with wavelengths spanning from 0.47 to 2.12
, adeptly employed for retrieving vital aerosol characteristics. In further detail, this instrumental suite produces daily-level aerosol optical thickness data, expertly mapped at a global spatial resolution of 10 km × 10 km. The MODIS swath width is 2330 km, slightly narrower than that of AVHRR. As a result, the coverage for a single day is not complete, and any gaps from one day are filled in on the next. For an animation illustrating the MODIS scan pattern, refer to http://aqua.nasa.gov/sites/default/files/aqua_modis_h264.mov (last accessed: 27 July 2023). The retrieval of aerosol data through MODIS employs three distinct algorithms, each designed for specific settings: DT and DB algorithms are employed over land, while the DT algorithm is used over oceans [64]. Additionally, the DTB algorithm harmonizes these main approaches, selecting the most suitable one based on the land’s characteristics. The DT algorithm is tailored for dark surfaces like dark soil and vegetation. Meanwhile, the advanced second-generation DB algorithm is adept at bright surfaces such as deserts, urban areas, and vegetated regions. The retrieval process of aerosol properties using the DT algorithm over dense vegetation and dark soil relies on how visible wavelengths, specifically 0.47 and 0.65
, correlate with a shortwave of 2.12
in the infra-red range [65]. The present operational MODIS dataset, denoted as C061, furnishes standard aerosol characteristics with a spatial resolution of 10 × 10 km2 within the Level 2 (L2) datasets, specifically MOD04 for Terra and MYD04 for Aqua. In aggregated Level 3 (L3) products, this resolution is lowered to 1° × 1°. Furthermore, an additional aerosol file based on the Dark Target (DT) approach with a resolution of 3 km is included in the C006 dataset, which has been continued in the current C061. This enhancement serves to offer air quality insights at local or urban scales. Detailed information about the C006 dataset is elaborated by [66], while the progression from the DT C006 to the current C061 is detailed from [67]. Although MODIS AOD products are widely used for aerosol monitoring due to their broad spatial coverage, they are not without limitations. Retrieval accuracy can be significantly affected by persistent cloud cover, haze, or high surface reflectance—conditions that are common in tropical regions like Ghana. The underlying retrieval algorithms (Dark Target and Deep Blue) rely on assumptions about surface reflectance and atmospheric conditions, which can introduce errors over bright or vegetated surfaces. Additionally, spatial and temporal coverage may be uneven due to the exclusion of low-quality retrievals under such conditions [68]. The daily AOD dataset at a wavelength of λ = 0.55
for Terra at 10:30 am local time and for Aqua at 1:30 pm local time during overpass times was downloaded from https://giovanni.gsfc.nasa.gov/giovanni/ (last accessed on 27 July 2023).
Input variables.
To predict the AOD from MODIS, climate variables such as temperature(t2m; K), dewpoint(K), surface net downward shortwave flux(), surface upward longwave flux (
), surface upward latent heat flux (
), relative humidity, boundary layer height(m), Downward UV radiation (
), Evaporation (m), Precipitation(m), Pressure (hPa), Top net solar radiation (
), and low cloud cover. We utilized these covariates, primarily relying on expert knowledge and data availability. Meteorological variables are known for their influence on the dispersion and transport of fine particulate matter. Atmospheric heat flux variables are also known to be influenced by aerosols which can lead to negative or positive radiative forcing. A summary table describing the physical relevance of each variable, its expected relationship with AOD, and supporting references is provided in Table A in S1 Text.
As previously mentioned, AOD is a key predictor of PM2.5, with higher levels of AOD indicating higher PM2.5 levels [35,40,69–71]. In general, the MinMaxScaler normalizer and the normalization layer from Keras tensorflow were used to normalize and standardize the input data. All climate variables used in prediction were obtained from ERA5 (https://cds.climate.copernicus.eu/) with a temporal range spanning 2003 to 2019. A total of 80% of the data was used for training and 20% was used for testing. The DT and deep blue DB combined products were used for this study. The decision to use this product was mainly due to the highly variable topography of our region. Here, we incorporate meteorological data to predict AOD levels, from 1st January 2003 to 31st December 2019 over the entire country and some selected locales.
Table 1 represents the statistical distribution of the input variables for the model.
Data preprocessing.
The original MODIS AOD datasets were obtained at a daily time resolution; however, considerable spatial and temporal gaps were observed in the data, particularly over terrestrial regions and during cloudy or hazy conditions. To address this challenge and ensure a cohesive analysis, we employed a methodology outlined by [37], which involved partitioning the country into smaller grids and averaging valid pixel values within each grid cell. A minimum of three observations was required to compute a representative value, reducing the influence of missing data while preserving meaningful spatial variability. To further improve data continuity and model stability, the daily AOD values were aggregated to a monthly time step. This regridding process significantly reduced the prevalence of missing values while retaining seasonal and interannual aerosol variability. During this process, 25 missing time steps were identified in the TERRA dataset. A complete list of missing time steps identified in the TERRA dataset is provided in the Table B in S1 Text. For consistency, the AQUA data were also adjusted to match this revised temporal structure. The list of these excluded datetimes is provided in the Table B in S1 Text. Grid cells with persistent gaps or retrievals flagged as low-quality in the QA layers were excluded. In total, 204 valid monthly observations were retained for each AOD product (AQUA and TERRA), as indicated in Table 1.
In this study, the entire country and specific major localities were considered, with AOD datasets from AQUA and TERRA serving as dependent variables, while other satellite-borne data products, including temperature, dewpoint, surface upward longwave flux, surface upward latent heat flux, surface net downward shortwave flux, relative humidity, boundary layer height, downward uv radiation, evaporation, precipitation, pressure, top net solar radiation, and low cloud cover, constituted the independent variables. Given the inherent variation in spatial resolution between the various data products utilized in this study, all data products not conforming to the 1 km × 1 km grid were standardized by bilinear interpolation. The bilinear interpolation method was chosen primarily for its versatility, as it strikes a balance between computational efficiency and accuracy [72]. This makes it suitable for a broad spectrum of applications where smooth interpolation between gridded data points is essential. Furthermore, the temporal resolution of the independent variables was adjusted to align with the Aqua and Terra AOD data sets, ensuring consistency.
Model development
One of the main goals of this study was to use ML to assess AOD over the entire country and some selected areas. As the no free lunch (NFL) theorem states; there is no single algorithm that performs best for all possible problems. Hence, we employed two well-known machine learning methods that are often used to predict AOD based on previous research [73].
Multiple linear regression(MLR).
The MLR model is a well-established machine learning algorithm frequently employed in studies focused on establishing linear relationships between multiple independent variables and a continuous dependent variable. Based on an extensive review of existing literature [73], it was found that the MLR model is the most commonly utilized in AOD machine learning studies. This popularity stems from its simplicity and effectiveness in handling linear correlations. Unlike certain other machine learning algorithms, the MLR model does not require hyperparameter tuning. However, as with many machine learning techniques, proper normalization of covariates is critical to ensuring optimal performance. In this study, the MinMaxScaler normalizer was used to standardize covariates, a choice motivated by its widespread use in scientific research for its efficiency and computational simplicity. The mathematical representation of the MLR model is provided in Equation A in S1 Text.
Multi linear regression (MLR) using ANN.
Before constructing the neural network using the optimal parameters determined through the grid search process, we employed the Artificial Neural Network (ANN) model for conducting a multi-linear regression analysis. This was executed by defining a single input layer and an output layer while excluding any hidden layers during model construction. The fundamental principle of linear regression is to capture linear associations between the dependent variable and independent variables. By structuring the ANN model with solely an input and output layer, it effectively operates akin to a linear regression model. This is mainly because the activation function is one of the fundamental differences between a linear regression model and a neural network, even with just input and output layers. Linear regression models directly output a weighted sum of inputs without applying any non-linear transformation. In contrast, neural networks typically use activation functions, even in the hidden layers, to introduce non-linearity. This non-linearity allows neural networks to breakdown complex patterns and relationships in the data. To assess the sensitivity of the model to different hyperparameter configurations, various combinations of epochs, batch sizes, and learning rates were evaluated using a trial-and-error approach. The number of epochs ranged from 50 to 200, with batch sizes varying between 8 and 24. Learning rates were tested at three different levels: 0.001, 0.01, and 0.1. Based on these experiments, the optimal configuration was identified as a batch size of 12, 89 epochs, and a learning rate of 0.01. The Adam optimizer was utilized for this model.
While hybrid MLR-ANN models have been explored in various domains, such as water quality prediction [74], chlorophyll-a concentration estimation [75], and crop yield forecasting [76], these approaches typically involve combining MLR and ANN models sequentially or in parallel to capture both linear and nonlinear relationships. To the best of our knowledge, structurally constraining an ANN to mimic MLR by eliminating hidden layers and using linear activation functions has not been previously applied in the context of AOD modeling in West Africa. Table 2 presents a summary of the performance metrics corresponding to some of the most effective hyperparameter combinations tested.
Artificial neural network (ANN).
ANNs are computational models inspired by the intricate neural networks in living organisms. They are renowned for their adept utilization of the backpropagation error technique, a pivotal method for training these networks [77]. This technique revolves around iteratively fine-tuning the network’s weights to minimize the discrepancy between anticipated and actual outputs. This intricate process encompasses several essential steps. In the forward pass, input data traverses through the network layer by layer. Neurons calculate weighted sums of inputs, subsequently processed by activation functions to yield neuron outputs. The distinction between predicted and target outputs is computed via a designated error or loss function, like the Mean Squared Error (MSE) for regression or Cross-Entropy for classification. Calculated derivatives guide the adjustment of weights and biases within the network using optimization algorithms, often Gradient Descent or its variants. The iterative repetition of this process over multiple epochs refines weights, progressively diminishing errors. Convergence towards an optimal weight configuration, minimizing the error function, characterizes the network’s learning trajectory. Influential hyperparameters like activation functions and learning rates can be selected through manual methods, involving trial and error, or alternative strategies such as grid search, random search, or Bayesian optimization. In this study, grid search was employed to identify optimal parameters, as detailed in Table 3. The training of the model involved the application of the Adam optimizer, primarily chosen mainly due to its fast convergence [78]. Additionally, the MSE was employed as the loss function during the training process. The mathematical representation of the MLPN model is provided in Equation B in S1 Text.
Hyper-parameter tuning.
While machine learning methods eliminate the need for certain distributional assumptions, they introduce a new challenge: hyperparameters. These are settings that influence the model’s learning process and must be carefully tuned. To optimize our model, we performed a grid search to fine-tune these hyperparameters, evaluating performance based on mean squared error and cross-validation values. For our neural network model, we further enhanced its performance by optimizing the architecture. This involved selecting three hidden layers, adjusting the number of neurons per layer, determining the optimal number of training iterations (epochs), and choosing suitable activation functions, batch sizes, and dropout rates. Additionally, we fine-tuned the learning rate to ensure the model adapted efficiently. These adjustments were made to ensure the model achieved the best possible performance. This whole tuning process was done using Keras tuner. Table 3 depicts the result of the tuning process.
Validation metrics
The control parameters of the models were initially chosen and then adjusted through trials to achieve the most optimal fitness measures. To assess the effectiveness of the proposed models, four statistical indicators were employed: root mean square error (RMSE) (Table A in S1 Text), coefficient of determination () (Equation D in S1 Text), mean absolute error (MAE) (Equation E in S1 Text), and the Kling-Gupta Efficiency (KGE) (Equation F in S1 Text). A detailed description of formulations are provided in the S1 Text.
Results and discussion
AOD distribution over Ghana
The climatological spatial distribution of AOD levels across Ghana is displayed in Figs 2 and 3. Over Ghana, the mean AOD values for Aqua and Terra were 0.477 (± 0.206 sd) and 0.504 (± 0.208 sd), respectively. The difference we observed in Aqua and Terra AOD retrievals in our study was approximately 6%. This difference, although noteworthy, contrasts the findings of [79] and [37], who independently reported a more substantial statistical difference of around 13% in global Aqua and Terra AOD retrievals. It is important to highlight that the discrepancy we observed, which is nearly half of their reported difference, may be linked to the fact that we integrated the DTB comprising the DT and DB products, as opposed to their utilization of stand-alone products.
Base map shapefile from OCHA via humanitarian data exchange (https://data.humdata.org/dataset/cod-ab-gha), licensed under CC BY 4.0.
Base map shapefile from OCHA via humanitarian data exchange (https://data.humdata.org/dataset/cod-ab-gha), licensed under CC BY 4.0.
Our observations reveal a concentration of elevated AODs in the southwestern region of the country, while sporadically elevated AOD fluctuations manifest along the mid and southeastern boundaries. These findings align with findings from [80] and [37]. [80] attributed their findings to the geographical characteristics of the southernmost regions of southern West Africa. They pointed out that these areas are predominantly characterized by low elevations. They observed that the multi-year averaged AOD550 (AOD at 550 nm) over this region exhibits an upward trend as the elevation decreases. This phenomenon can be explained by the influence of high terrain, which can either impede or modify wind directions. Such alterations in wind patterns disrupt the horizontal dispersion of pollutants [81,82], consequently leading to the diminishing in pollutant concentration levels. Furthermore, [37] provided additional insights into the factors influencing the elevated AOD values observed in southwestern Ghana. These heightened AOD values are postulated to arise from various sources, including the presence of sea salt spray suspended within the atmospheric column and aerosols originating from specific source regions. These components not only directly contribute to localized dust accumulations but also exert a significant influence on overall aerosol loadings. In contrast, the comparatively higher AODs detected in the middle and eastern sectors of Ghana are likely linked to anthropogenic activities. These activities encompass emissions of both fine and coarse PM, primarily associated with surface mining operations. Furthermore, these elevated AOD levels may result from emissions of BC originating from biomass combustion and aerosols transported from distant source regions. The complex interplay of these factors, coupled with the unique regional topography, contributes to the observed AOD patterns. Of particular note is the presence of the Akuapim-Togo mountain range within the eastern corridors of Ghana. This geographical feature significantly influences the dispersion and containment of aerosols within the region. Consequently, the topographical characteristics of the elevated terrain likely play a crucial role in contributing to the relatively higher AOD levels observed in the eastern sector of the country.
One potential contributor to elevated AOD levels in parts of southwestern Ghana can also be attributed to the prevalence of illegal timber harvesting activities. Studies estimate that up to 70% of Ghana’s total timber harvest is illegal, often involving unsupervised chainsaw operations at night that lead to deforestation, canopy degradation, and the accumulation of plant debris and coarse particulate matter [23,24,83]. These activities are concentrated in forest reserves such as Atewa, Boin Tano, Asenanyo, and Apedwa, many of which are located near urban centers like Takoradi and Kumasi. The associated aerosol emissions, particularly during dry seasons, may partially explain elevated AOD levels in these regions. Although the MODIS AOD dataset used in this study (∼10 km resolution) may not resolve localized forest disturbances, it is suitable for detecting broad regional patterns. Future work could integrate higher-resolution satellite imagery (e.g., Sentinel-2 or Landsat) and apply unsupervised classification techniques within a GIS framework to better characterize land-use changes and their contributions to aerosol loading.
It is also worth noting that several AOD hotspots identified in the southwestern region—including Tarkwa Nsuaem, Prestea Huni-Valley, and Wassa Amenfi East—spatially overlap with areas of intense illegal surface mining activity (galamsey). These operations, which disturb large areas of land and generate significant dust emissions, have been shown to concentrate heavily in these districts, with over 7,400 sites mapped in a study by [84]. This likely contributes to the persistently high aerosol levels observed over these regions. These observations highlight the complex interrelationships between geographical features, human activities, and aerosol dynamics in shaping regional aerosol distribution patterns. Seasonal patterns of AOD (Aqua and Terra) are presented in the Fig B and Fig C in S1 Text to illustrate intra-annual variability across Ghana.
Feature importance
According to [85] and [86], the process of feature selection stands out as a crucial phase in machine learning, possibly even more important than the model selection itself. Understanding the individual contribution of each covariate to the model’s performance is indispensable. The inclusion of irrelevant variables and highly correlated variables, often referred to as multi-collinearity, can significantly impair a model’s effectiveness. In this study, we employed both correlation analysis and the intrinsic feature importance attribute derived from the random forest (RF) algorithm. These methods were instrumental in our quest to identify the most optimal features for our models. Fig 4 provides a comprehensive visualization of the contributions made by various covariates for both the AQUA and TERRA datasets, obtained through the RF feature importance property. For TERRA, the most influential feature was the downward shortwave flux, contributing approximately 50%, followed by the latent heat flux, which accounted for around 12% of the variance. Notably, the other features contributed less than 10% individually. Conversely, in the case of AQUA, the downward shortwave flux remained paramount, contributing about 45%. Here, the influence of latent heat flux decreased significantly, from 12% to 8%. Intriguingly, the contribution of sensible heat flux rose notably from 8% to 17%, making it the second most influential feature for predicting AQUA. The shift between sensible and latent heat flux concerning AQUA and TERRA can be attributed to the variance in satellite overpass times. For TERRA, the overpass transpires in the morning when the sun has yet to reach its zenith. Overnight, surface temperatures tend to decrease, causing any moisture present on surfaces to undergo a phase change, releasing latent heat. This latent heat energy profoundly influences the aerosols detected by the satellite during its morning overpass. Conversely, AQUA’s overpass occurs in the afternoon when the sun shines brightly. By this time, the sun has sufficiently warmed the surface, causing the adjacent air to heat up. Consequently, this warm air ascends, carrying heat energy with it, thereby influencing the aerosols detected by the satellite. These findings align with the established understanding of aerosol impacts on energy exchange processes in the atmosphere, as supported by previous studies [87–90], emphasizing the dynamic interplay between aerosols and surface-atmosphere energy fluxes. Fig 5 presents the annual mean AOD values, the number of days per year with high AOD (defined as AOD>0.509, the national average), and monthly climatological patterns. These results suggest that while satellite overpass timing may influence model sensitivity to different physical drivers (e.g., latent vs. sensible heat flux), it does not necessarily translate to a consistent difference in AOD magnitude between AQUA and TERRA across all time periods. Our analysis shows that, in practice, TERRA AOD values are often higher, which contrasts with initial expectations based on afternoon convective mixing arguments. This pattern is consistent with previous findings over Ghana by [37].
Variables include shortwave radiation (shortwave), latent heat flux (latent), sensible heat flux (sensible), evaporation rate (evaporation), surface temperature (temp), surface pressure (pressure), precipitation (precip), low cloud cover (lcc), solar radiation (solar_radiation), boundary layer height (blh), dew point temperature (dewpoint), surface longwave radiation (longwave), relative humidity (RH), and ultraviolet radiation (uv_radiation). Results highlight that shortwave radiation and flux-based variables dominate AOD prediction, with notable differences in feature ranking between morning (TERRA) and afternoon (AQUA) satellite overpasses.
Top left panel: annual mean AOD. Top right panel: monthly climatology Bottom: number of high AOD days (AOD>0.509 (national average)).
Fig 6 illustrates the correlation matrix, a fundamental analytical tool employed in this study. The correlation matrix serves a dual purpose: first, it aids in identifying covariates that exhibit strong correlations with the target variable. Secondly, it facilitates the detection of variables displaying high intercorrelations, thereby assisting in mitigating the challenge of multicollinearity.
Variables include boundary layer height (blh), sensible heat flux (sensible), latent heat flux (latent), surface temperature (temp), low cloud cover (lcc), net shortwave radiation (shortwave), dew point temperature (dewpoint), surface longwave radiation (longwave), relative humidity (RH), evaporation, surface pressure (pressure), total solar radiation (solar radiation), ultraviolet radiation (uv radiation), and precipitation (precip). Positive correlations are shown in red and negative correlations in blue. Strong correlations (absolute r>0.6) highlight key variables potentially influencing AOD behavior across different times of day and surface conditions.
The models were trained using eight climate variables selected for their relevance to surface-atmosphere interactions that affect aerosol dynamics: downward shortwave flux, latent heat flux, sensible heat flux, evaporation, surface pressure, temperature and boundary layer height (BLH). While some of these, such as shortwave and latent heat flux, may themselves be influenced by aerosol concentrations, they also capture atmospheric energy partitioning and moisture availability that co-vary with aerosol processes. For example, sensible and latent heat fluxes are shaped by boundary layer dynamics, land surface processes, and vegetation physiology, which are in turn modulated by atmospheric thermodynamics. Including variables such as pressure, temperature, and BLH allows the model to learn these overlapping signals, using both aerosol-induced and meteorologically driven variability. Their inclusion is based on established empirical links in the literature (e.g., [87–89]) and on the objective of using a wide range of variables to predict AOD in the absence of ground-based aerosol data. Importantly, the model is designed to uncover statistical associations rather than infer causality.
Spatial prediction performance and localized assessment of predicted AOD levels
To evaluate the ability of the machine learning (ML) models to predict the spatio-temporal distribution of AOD at a resolution of 1 km, we applied a suite of performance metrics: RMSE, MAE, , and KGE, across the study region (see section “Validation metrics”). The models were trained and evaluated separately for AQUA and TERRA datasets, with Figs 7 and 8 visualizing spatial model performance across the country.
Base map shapefile from OCHA via humanitarian data exchange (https://data.humdata.org/dataset/cod-ab-gha), licensed under CC BY 4.0.
Base map shapefile from OCHA via humanitarian data exchange (https://data.humdata.org/dataset/cod-ab-gha), licensed under CC BY 4.0.
All models demonstrated a strong predictive skill in the northern and middle sectors, with values of RMSE and MAE ranging from 0.025 to 0.175 and values exceeding 0.5. However, performance continued to decline in the southwest region, where RMSE values approached 0.2 and KGE dropped below 0.3 in some grid cells. This subtle performance is likely related to the complex land-atmosphere interactions in this region, including dense vegetation cover and unobserved anthropogenic activities such as nighttime timber harvesting ([23,24,83]), which introduce aerosol variability that cannot be captured using meteorological predictors alone. As noted in [91] and [37], such areas may produce significant biogenic aerosols, complicating the estimation of AOD. The lack of ground-based observations also limited the inclusion of such emissions datasets or land-use activity, which could have enhanced model learning in these regions.
The spatial patterns of AOD in Ghana (see section “AOD distribution over Ghana”), together with the regionally varying performance of the machine learning models -particularly in areas influenced by complex land-atmosphere interactions - informed the selection of two major urban hubs, Accra and Takoradi, for further localized evaluation.
These cities were identified as aerosol hotspots, with mean AOD values consistently exceeding the national average of 0.509. Accra recorded mean AODs of 0.703 (TERRA) and 0.587 (AQUA), while Takoradi recorded 0.701 (TERRA) and 0.622 (AQUA), as shown in Fig 9. These elevated levels are particularly relevant for public health, as long-term exposure to elevated levels of fine particulate matter has been associated with a spectrum of health risks. These include aggravated respiratory symptoms such as asthma and chronic obstructive pulmonary disease (COPD), cardiovascular irregularities such as vascular dysfunction and arrhythmia, and even an increased risk of respiratory cancers and premature mortality [31].
The causes of aerosol burden in these locales differ: Accra levels are likely driven by heavy vehicular traffic, industrial emissions, and population growth, while Takoradi levels are influenced by offshore oil and gas activities and coastal pollution transport [92,93]. These environmental complexities provided an opportunity to evaluate the generalizability of the models under different anthropogenic pressures.
Fig 10 presents a comparison of observed and predicted AOD values in both cities. The ANN(MLR) model consistently outperformed MLR and traditional ANN models, with RMSEs ranging from 0.11 to 0.18 and KGE values reaching 0.76 in Takoradi. In Accra, ANN(MLR) achieved the highest and KGE, followed closely by MLR. Although traditional ANN models showed competitive MAE values, they lagged in overall agreement metrics such as KGE. A detailed summary of the various metrics employed in this analysis is presented in Table 4.
These results demonstrate that, despite data limitations, ML models, particularly ANN (MLR), can simulate regional AOD dynamics with high fidelity. The ability to localize model predictions to densely populated areas highlights their potential for air quality surveillance and health risk assessment in data-sparse environments. As such, AOD modeling, supported by satellite products and ML techniques, provides a valuable proxy for assessing chronic exposure to particulate pollution in rapidly urbanizing settings.
Limitations
Although this study demonstrates the potential of machine learning models to simulate AOD variability in data-scarce regions, there are some limitations that should be acknowledged.
- The MODIS AOD dataset used in this study, although globally recognized and validated, has a spatial resolution of 1 km, which may be insufficient to capture fine-scale variations in aerosol concentrations, especially in dense urban or industrial environments.
- Only meteorological and surface energy balance variables were used as predictors, without the inclusion of critical variables such as land use data, emission inventories, vegetation cover, or traffic-related activity. This may limit the ability of the models to fully capture the drivers of aerosol variability.
- The lack of a dense ground-based monitoring network across Ghana meant that model outputs could not be validated against in situ PM2.5 or AOD measurements, restricting the evaluation of model accuracy and limiting the linkage between modeled AOD and actual surface-level exposure.
- Persistent missing values, especially in the daily MODIS AOD records, led to the removal of 25 TERRA time steps and required re-gridding to a monthly resolution. This improved spatial coverage, but reduced the temporal richness of the dataset and may have smoothed out short-term aerosol events.
Conclusions
The study delineates a comprehensive spatio-temporal assessment of aerosol distribution across Ghana and two of its prominent cities using MODIS AOD data at a spatial resolution of 1 km. This investigation spans a 16-year period (2003-2019) and delves into intricate patterns of aerosol distribution and concentration. AOD is a key predictor of PM2.5, with higher levels of AOD indicating higher PM2.5 levels, a relationship highlightd by several referenced studies.
Guided by the principle of the No Free Lunch (NFL) theorem, which highlights the absence of a universally optimal algorithm for all problems, we evaluated the performance of two distinct machine learning algorithms in predicting AOD values throughout the country and some selected locales, specifically Accra and Takoradi. The selection of these regions is based on their status as the major cities of Ghana, characterized by the highest mean aerosol burden compared to other important urban centers. To the best of our knowledge, this study is the first to use machine learning models to perform AOD assessments over the country. The following conclusions are drawn:
- The analysis of spatio-temporal aerosol distribution revealed noteworthy insights. The examination of MODIS Aqua and Terra AOD retrievals unveiled an overall relatively lower aerosol burden over Ghana, marked by mean AOD values hovering around 0.509. Moreover, the retrieval patterns demonstrated a subtle variance of approximately 0.07 between the mean Terra and Aqua AODs. Our research reveals distinctive patterns in aerosol concentration across various regions within our study area. The southwestern part of the country consistently exhibits elevated aerosol loadings, while the northern, eastern coastal areas and some parts of the middle sector, consistently display lower aerosol concentrations. This recurring observation aligns with prior studies undertaken in the same geographical vicinity [37,80] and can be ascribed to various factors. The presence of dense vegetation in the southwestern region likely contributes to this aerosol distribution pattern, possibly linked to increased biogenic emissions. Additionally, [37] suggested that elevated aerosol loadings in the southwestern sector of the country result from a combination of factors, including the complex dynamics of sea salt spray deposition from oceanic bubble eruptions and emissions from the petroleum and gas sectors along the western coast. The proximity of these coastal phenomena to the southwestern region significantly contributes to higher aerosol concentrations. [80] also attributed their findings to geographical characteristics in southern West Africa, where low elevations prevail and elevated terrain can alter wind patterns, affecting pollutant dispersion. Additionally, persistently elevated AOD levels observed in the southwestern region appear to coincide with areas of intensive illegal surface mining (galamsey) as reported by [84], suggesting that localized land disturbance and dust emissions from these operations may be significant contributors to regional aerosol burdens.
- Utilizing a comprehensive range of validation metrics for assessing model performance, we can confidently conclude all models developed in the study exhibited an acceptable level of accuracy. However, the MLR utilizing the ANN architecture ANN(MLR), exhibited superior predictive capabilities compared to both the classic MLR and the standard Keras-tuned ANN model.
- The sub-optimal performance of the standard ANN model is consistent with findings from prior research [94,95], which have shown that feed-forward neural networks often struggle with accuracy and saturation issues as the number of hidden layers increases. These challenges are commonly observed in similar studies and can be linked to several inherent limitations of the model. Overfitting is a primary concern, where the model becomes overly specialized to the training data, resulting in poor generalization on new, unseen data. Another limitation is the model’s need for substantial amounts of data, which can be a barrier in cases with smaller datasets. Additionally, the ANN is susceptible to local minima traps and the exploding gradient problem, both of which can hinder the model’s performance. Further details on these limitations can be found in [96].
- Elevated pollution levels, driven by both biogenic and anthropogenic emissions, are a defining feature of the selected major cities. However, during the preprocessing of MODIS data, a significant presence of NaN (Not a Number) values was detected. This issue, combined with limited ground-level observations, significantly hampers our ability to monitor and forecast air quality in these urban areas and across the wider region. Machine learning models offer a promising solution to address these challenges. By utilizing easily accessible data, these models provide a reliable method for predicting and monitoring air quality with reasonable accuracy, despite the limitations posed by data gaps, thereby enabling more comprehensive environmental assessments.
- For future studies, it’s worth noting that the datasets we used in our research have a relatively basic level of detail in terms of spatial resolution. To improve the precision of upcoming studies, it would be beneficial to consider datasets from other data sources with finer spatial resolutions and longer duration (∼30 or more years). These more detailed datasets would allow for quicker and more thorough assessments of aerosol distributions and related factors, providing a more comprehensive understanding of the subject matter. Also, our analysis primarily concentrated on meteorological variables as input data for the models. However, it is recommended that future investigations consider the integration of additional variables, such as land use characteristics, population growth patterns, vehicular emissions, agricultural residue burning, domestic waste burning, industrial/biogenic emissions, the density of transportation hubs, and daily traffic counts. These supplementary variables can provide a more holistic understanding of the factors influencing aerosol distribution and facilitate more comprehensive predictive models.
Code availability
All codes used for data preprocessing, machine learning model development, and figure generation are available in a public GitHub repository: https://github.com/hefe2020/AOD_ML_Project.
Supporting information
S1 Text.
Supplementary Information. Contains:
https://doi.org/10.1371/journal.pclm.0000651.s001
- – Table A: Physical relevance of meteorological variables used in AOD prediction.
- – Table B: Missing time steps identified in the TERRA dataset.
- – Equations A–F: Mathematical formulations for regression, ANN, and evaluation metrics (RMSE, MAE, R2, KGE).
- – Fig A: Architecture of the Artificial Neural Network (ANN).
- – Fig B: Seasonal distribution of AOD from AQUA.
- – Fig C: Seasonal distribution of AOD from TERRA.
References
- 1.
Kondratyev K, Ivlev L, Krapivin V, Varostos C. Atmospheric aerosol properties: Formation, processes and impacts. Springer Science & Business Media; 2006. https://doi.org/10.1007/3-540-37698-4
- 2. Putaud J-P, Van Dingenen R, Alastuey A, Bauer H, Birmili W, Cyrys J, et al. A European aerosol phenomenology – 3: Physical and chemical characteristics of particulate matter from 60 rural, urban, and kerbside sites across Europe. Atmos Environ. 2010;44(10):1308–20.
- 3.
Boucher O, Randall D, Artaxo P, Bretherton C, Feingold G, Forster P. Clouds and aerosols. Cambridge University Press; 2013. p. 571–657.
- 4. Yu H, Kaufman YJ, Chin M, Feingold G, Remer LA, Anderson TL, et al. A review of measurement-based assessments of the aerosol direct radiative effect and forcing. Atmos Chem Phys. 2006;6(3):613–66.
- 5. Lohmann U, Feichter J. Global indirect aerosol effects: A review. Atmos Chem Phys. 2005;5(3):715–37.
- 6. Johnson BT, Shine KP, Forster PM. The semi-direct aerosol effect: Impact of absorbing aerosols on marine stratocumulus. Quart J Royal Meteoro Soc. 2004;130(599):1407–22.
- 7. Creamean JM, Suski KJ, Rosenfeld D, Cazorla A, DeMott PJ, Sullivan RC, et al. Dust, biological aerosols from the Sahara and Asia influence precipitation in the western US. Science. 2013 ;339(6127):1572–8.
- 8. Misra AK. Climate change and challenges of water and food security. Int J Sustain Built Environ. 2014;3(1):153–65.
- 9.
World Health Organization WHO. Health effects of particulate matter: Policy implications for countries in Eastern Europe, Caucasus and Central Asia. World Health Organization; 2013.
- 10.
Kato Y. Application of dust and PM2.5 detection methods using MODIS data to the Asian dust events which aggravated respiratory symptoms in Western Japan in May 2011. In: Remote Sensing of the Atmosphere, Clouds, and Precipitation VII; 2018. p. 57. https://doi.org/10.1117/12.2324735
- 11. Prieto-Parra L, Yohannessen K, Brea C, Vidal D, Ubilla CA, Ruiz-Rudolph P. Air pollution, PM2.5 composition, source factors, and respiratory symptoms in asthmatic and nonasthmatic children in Santiago, Chile. Environ Int. 2017;101:190–200. pmid:28202226
- 12. Zeng X, Xu X, Zheng X, Reponen T, Chen A, Huo X. Heavy metals in PM2.5 and in blood, and children’s respiratory symptoms and asthma from an e-waste recycling area. Environ Pollut. 2016;210:346–53. pmid:26803791
- 13. Jung C-R, Young L-H, Hsu H-T, Lin M-Y, Chen Y-C, Hwang B-F, et al. PM2.5 components and outpatient visits for asthma: A time-stratified case-crossover study in a suburban area. Environ Pollut. 2017;231(Pt 1):1085–92. pmid:28922715
- 14. Williams AP, Abatzoglou JT, Gershunov A, Guzman-Morales J, Bishop DA, Balch JK, et al. Observed impacts of anthropogenic climate change on wildfire in California. Earth’s Future. 2019;7(8):892–910.
- 15. Dabass A, Talbott EO, Rager JR, Marsh GM, Venkat A, Holguin F, et al. Systemic inflammatory markers associated with cardiovascular disease and acute and chronic exposure to fine particulate matter air pollution (PM2.5) among US NHANES adults with metabolic syndrome. Environ Res. 2018;161:485–91. pmid:29223110
- 16. Vidale S, Campana C. Ambient air pollution and cardiovascular diseases: From bench to bedside. Eur J Prev Cardiol. 2018;25(8):818–25. pmid:29580063
- 17. Thaller EI, Petronella SA, Hochman D, Howard S, Chhikara RS, Brooks EG. Moderate increases in ambient PM2.5 and ozone are associated with lung function decreases in beach lifeguards. J Occup Environ Med. 2008;50(2):202–11. pmid:18301177
- 18. Apte JS, Marshall JD, Cohen AJ, Brauer M. Addressing global mortality from ambient PM2.5. Environ Sci Technol. 2015;49(13):8057–66. pmid:26077815
- 19. David LM, Ravishankara AR, Kodros JK, Pierce JR, Venkataraman C, Sadavarte P. Premature mortality due to PM2.5 over India: Effect of atmospheric transport and anthropogenic emissions. Geohealth. 2019;3(1):2–10. pmid:32159019
- 20.
Ministry of Transport. Ministry of transport; 2016. https://en.wikipedia.org/wiki/Ministry_of_Transport_(Ghana).
- 21. Sierra-Vargas MP, Teran LM. Air pollution: Impact and prevention. Respirology. 2012;17(7):1031–8. pmid:22726103
- 22. Cai S, Wang Y, Zhao B, Wang S, Chang X, Hao J. The impact of the “Air pollution prevention and control action plan” on PM2.5 concentrations in Jing-Jin-Ji region during 2012 -2020. Sci Total Environ. 2017;580:197–209. pmid:28011024
- 23.
Sarfo-Mensah P. Exportation of timber in Ghana: The menace of illegal logging operations. Econstor. 2005. https://hdl.handle.net/10419/74062
- 24. Boakye J. Estimation of illegal logging by the formal timber sector in Ghana: Implications for forest law compliance, enforcement and EU-Ghana voluntary partnership agreement. Int Forest Rev. 2015;17(2):117–27.
- 25. Wu Y, de Graaf M, Menenti M. Improved MODIS Dark Target aerosol optical depth algorithm over land: Angular effect correction. Atmos Meas Tech. 2016;9(11):5575–89.
- 26. Mhawish A, Banerjee T, Broday DM, Misra A, Tripathi SN. Evaluation of MODIS Collection 6 aerosol retrieval algorithms over Indo-Gangetic Plain: Implications of aerosols types and mass loading. Remote Sens Environ. 2017;201:297–313.
- 27. Shen X, Bilal M, Qiu Z, Sun D, Wang S, Zhu W. Validation of MODIS C6 dark target aerosol products at 3 km and 10 km spatial resolutions over the China seas and the eastern Indian ocean. Remote Sens. 2018;10(4):573.
- 28. Wei J, Peng Y, Guo J, Sun L. Performance of MODIS collection 6.1 Level 3 aerosol products in spatial-temporal variations over land. Atmos Environ. 2019;206:30–44.
- 29. Lee HJ, Liu Y, Coull BA, Schwartz J, Koutrakis P. A novel calibration approach of MODIS AOD data to predict PM2.5 concentrations. Atmos Chem Phys. 2011;11(15):7991–8002.
- 30. Dandou A, Bosioli E, Tombrou M, Sifakis N, Paronis D, Soulakellis N, et al. The importance of mixing height in characterising pollution levels from aerosol optical thickness derived by satellite. Water Air Soil Pollut: Focus. 2002;2(5–6):17–28.
- 31. Kim D, Kim J, Jeong J, Choi M. Estimation of health benefits from air quality improvement using the MODIS AOD dataset in Seoul, Korea. Environ Res. 2019;173:452–61. pmid:30978520
- 32. Kanabkaew T. Prediction of hourly particulate matter concentrations in Chiangmai, Thailand using MODIS aerosol optical depth and ground-based meteorological data. Environ Asia. 2013;6(2):65–70.
- 33. Hu X, Waller LA, Al-Hamdan MZ, Crosson WL, Estes MG Jr, Estes SM, et al. Estimating ground-level PM(2.5) concentrations in the southeastern U.S. using geographically weighted regression. Environ Res. 2013;121:1–10. pmid:23219612
- 34. Xu J, Jiang H, Xiao Z, Wang B, Wu J, Lv X. Estimating air particulate matter using MODIS data and analyzing its spatial and temporal pattern over the Yangtze Delta region. Sustainability. 2016;8(9):932.
- 35. Liu Y, Franklin M, Kahn R, Koutrakis P. Using aerosol optical thickness to predict ground-level PM2.5 concentrations in the St. Louis area: A comparison between MISR and MODIS. Remote Sens Environ. 2007;107(1–2):33–44.
- 36. Kahn RA, Gaitley BJ. An analysis of global aerosol type as retrieved by MISR. JGR Atmos. 2015;120(9):4248–81.
- 37. Fosu-Amankwah K, Bessardon GEQ, Quansah E, Amekudzi LK, Brooks BJ, Damoah R. Assessment of aerosol burden over Ghana. Sci Afr. 2021;14:e00971.
- 38. Engel-Cox JA, Holloman CH, Coutant BW, Hoff RM. Qualitative and quantitative evaluation of MODIS satellite sensor data for regional and urban scale air quality. Atmos Environ. 2004;38(16):2495–509.
- 39. Chu DA, Kaufman YJ, Zibordi G, Chern JD, Mao J, Li C, et al. Global monitoring of air pollution over land from the earth observing system-terra moderate resolution imaging spectroradiometer (MODIS). J Geophys Res. 2003;108(D21).
- 40. Gupta P, Christopher SA, Wang J, Gehrig R, Lee Y, Kumar N. Satellite remote sensing of particulate matter and air quality assessment over global cities. Atmos Environ. 2006;40(30):5880–92.
- 41. Kloog I, Nordio F, Coull BA, Schwartz J. Incorporating local land use regression and satellite aerosol optical depth in a hybrid model of spatiotemporal PM2.5 exposures in the Mid-Atlantic states. Environ Sci Technol. 2012;46(21):11913–21. pmid:23013112
- 42. Geng G, Zhang Q, Martin RV, van Donkelaar A, Huo H, Che H, et al. Estimating long-term PM2.5 concentrations in China using satellite-based aerosol optical depth and a chemical transport model. Remote Sens Environ. 2015;166:262–70.
- 43. Danesh Yazdi M, Kuang Z, Dimakopoulou K, Barratt B, Suel E, Amini H, et al. Predicting fine particulate matter (PM2.5) in the greater London area: An ensemble approach using machine learning methods. Remote Sens. 2020;12(6):914.
- 44. Taheri Shahraiyni H, Sodoudi S. Statistical modeling approaches for PM10 prediction in urban areas; a review of 21st-century studies. Atmosphere. 2016;7(2):15.
- 45. Li L. A robust deep learning approach for spatiotemporal estimation of satellite AOD and PM2.5. Remote Sens. 2020;12(2):264.
- 46. Nabavi SO, Haimberger L, Abbasi E. Assessing PM2.5 concentrations in Tehran, Iran, from space using MAIAC, deep blue, and dark target AOD and machine learning algorithms. Atmos Pollut Res. 2019;10(3):889–903.
- 47. Bai Y, Zeng B, Li C, Zhang J. An ensemble long short-term memory neural network for hourly PM2.5 concentration forecasting. Chemosphere. 2019;222:286–94. pmid:30708163
- 48. Xiao F, Yang M, Fan H, Fan G, Al-Qaness MAA. An improved deep learning model for predicting daily PM2.5 concentration. Sci Rep. 2020;10(1):20988. pmid:33268885
- 49. Di Q, Koutrakis P, Schwartz J. A hybrid prediction model for PM2.5 mass and components using a chemical transport model and land use regression. Atmos Environ. 2016;131:390–9.
- 50. Taghavi-Shahri SM, Fassò A, Mahaki B, Amini H. Concurrent spatiotemporal daily land use regression modeling and missing data imputation of fine particulate matter using distributed space-time expectation maximization. Atmos Environ. 2020;224:117202.
- 51. Abokyi E, Appiah-Konadu P, Abokyi F, Oteng-Abayie EF. Industrial growth and emissions of CO2 in Ghana: The role of financial development and fossil fuel consumption. Energy Rep. 2019;5:1339–53.
- 52. Aryee JNA, Amekudzi LK, Quansah E, Klutse NAB, Atiah WA, Yorke C. Development of high spatial resolution rainfall data for Ghana. Int J Climatol. 2017;38(3):1201–15.
- 53. Reeves CE, Formenti P, Afif C, Ancellet G, Attié J-L, Bechara J, et al. Chemical and aerosol characterisation of the troposphere over West Africa during the monsoon period as part of AMMA. Atmos Chem Phys. 2010;10(16):7575–601.
- 54. De Longueville F, Hountondji Y-C, Henry S, Ozer P. What do we know about effects of desert dust on air quality and human health in West Africa compared to other regions?. Sci Total Environ. 2010;409(1):1–8. pmid:20934742
- 55. Amoatey P, Omidvarborna H, Baawain M. The modeling and health risk assessment of PM2.5 from Tema Oil Refinery. Human Ecol Risk Assess: Int J. 2017;24(5):1181–96.
- 56. Agyemang-Bonsu KW, Dontwi IK, Tutu-Benefoh DA, Bentil DE, Boateng OG, Asuobonteng K, et al. Traffic-data driven modelling of vehicular emissions using COPERT III in Ghana: A case study of Kumasi. Am J Sci Ind Res. 2010;1(1):32–40.
- 57.
World Health Organization. World Health Statistics 2016. World Health Organization; 2016. https://www.who.int/docs/default-source/gho-documents/world-health-statistic-reports/world-health-statistics-2016.pdf
- 58. Odonkor ST, Mahami T. Knowledge, attitudes, and perceptions of air pollution in Accra, Ghana: A critical survey. J Environ Public Health. 2020;2020:3657161. pmid:32104187
- 59. Barnes WL, Xiong X, Salomonson VV. Status of terra MODIS and aqua modis. Advances in Space Research. 2003;32(11):2099–106.
- 60. Kahn RA, Nelson DL, Garay MJ, Levy RC, Bull MA, Diner DJ, et al. MISR aerosol product attributes and statistical comparisons with MODIS. IEEE Trans Geosci Remote Sens. 2009;47(12):4095–114.
- 61. Zhang J, Reid JS. An analysis of clear sky and contextual biases using an operational over ocean MODIS aerosol product. Geophys Res Lett. 2009;36(15).
- 62. Shi K, Zhang Y, Zhu G, Liu X, Zhou Y, Xu H, et al. Long-term remote monitoring of total suspended matter concentration in Lake Taihu using 250m MODIS-Aqua data. Remote Sens Environ. 2015;164:43–56.
- 63. Acharya P, Sreekesh S. Seasonal variability in aerosol optical depth over India: A spatio-temporal analysis using the MODIS aerosol product. Int J Remote Sens. 2013;34(13):4832–49.
- 64. Hsu NC, Lee J, Sayer AM, Kim W, Bettenhausen C, Tsay S -C.. VIIRS deep blue aerosol products over land: Extending the EOS long-term aerosol data records. JGR Atmos. 2019;124(7):4026–53.
- 65. Levy RC, Remer LA, Mattoo S, Vermote EF, Kaufman YJ. Second-generation operational algorithm: Retrieval of aerosol properties over land from inversion of Moderate Resolution Imaging Spectroradiometer spectral reflectance. J Geophys Res. 2007;112(D13).
- 66. Remer LA, Mattoo S, Levy RC, Munchak LA. MODIS 3 km aerosol product: Algorithm and global perspective. Atmos Meas Tech. 2013;6(7):1829–44.
- 67.
Mattoo S. Aerosol dark target (10km & 3km) collection 6.1 changes; 2017. https://Atmosphere-imager.gsfc.nasa.gov/sites/default/files/ModAtmo/C061_Aerosol_Dark_Target_v2.pdf
- 68. Levy RC, Remer LA, Kleidman RG, Mattoo S, Ichoku C, Kahn R, et al. Global evaluation of the Collection 5 MODIS dark-target aerosol products over land. Atmos Chem Phys. 2010;10(21):10399–420.
- 69. Wang J, Liu X, Christopher SA, Reid JS, Reid E, Maring H. The effects of non-sphericity on geostationary satellite retrievals of dust aerosols. Geophys Res Lett. 2003;30(24).
- 70. van Donkelaar A, Martin RV, Park RJ. Estimating ground-level PM2.5 using aerosol optical depth determined from satellite remote sensing. J Geophys Res. 2006;111(D21).
- 71. van Donkelaar A, Martin RV, Brauer M, Kahn R, Levy R, Verduzco C, et al. Global estimates of ambient fine particulate matter concentrations from satellite-based aerosol optical depth: Development and application. Environ Health Perspect. 2010;118(6):847–55. pmid:20519161
- 72.
Arif F, Akbar M. Resampling air borne sensed data using bilinear interpolation algorithm. [Conference Paper]. 2005:62–5.
- 73. Chu Y, Liu Y, Li X, Liu Z, Lu H, Lu Y, et al. A review on predicting ground PM2.5 concentration using satellite aerosol optical depth. Atmosphere. 2016;7(10):129.
- 74. Boah DK, Twum SB. Application of a hybrid multiple linear regression (MLR) and artificial neural network (ANN) model for water quality prediction. Afr Scholar J Afr Sustain Dev. 2021;23(2):1–15.
- 75. Lola MS, Ramlee MNA, Gunalan GS, Zainuddin NH, Zakariya R, Idris M, et al. Improved the prediction of multiple linear regression model performance using the hybrid approach: A case study of chlorophyll-a at THE OFFSHORE KUALA Terengganu, Terengganu. OJS. 2016;06(05):789–804.
- 76. Maya Gopal PS, Bhargavi R. A novel approach for efficient crop yield prediction. Comput Electr Agric. 2019;165:104968.
- 77. Agatonovic-Kustrin S, Beresford R. Basic concepts of artificial neural network (ANN) modeling and its application in pharmaceutical research. J Pharm Biomed Anal. 2000;22(5):717–27. pmid:10815714
- 78.
Bock S, Goppold J, Weiß M. An improvement of the convergence proof of the ADAM-optimizer. arXiv preprint; 2018. https://doi.org/10.48550/arXiv.1804.10587
- 79. Levy RC, Mattoo S, Sawyer V, Shi Y, Colarco PR, Lyapustin AI, et al. Exploring systematic offsets between aerosol products from the two MODIS sensors. Atmos Meas Tech. 2018;11(7):4073–92. pmid:32676129
- 80. Aklesso M, Kumar KR, Bu L, Boiyo R. Analysis of spatial-temporal heterogeneity in remotely sensed aerosol properties observed during 2005 –2015 over three countries along the Gulf of Guinea Coast in Southern West Africa. Atmos Environ. 2018;182:313–24.
- 81. Ma Z, Hu X, Sayer AM, Levy R, Zhang Q, Xue Y, et al. Satellite-based spatiotemporal trends in PM2.5 concentrations: China 2004 -2013. Environ Health Perspect. 2016;124(2):184–92. pmid:26220256
- 82. Ning G, Wang S, Ma M, Ni C, Shang Z, Wang J, et al. Characteristics of air pollution in different zones of Sichuan Basin, China. Sci Total Environ. 2018;612:975–84. pmid:28892849
- 83. Hansen CP, Treue T. Assessing illegal logging in Ghana. Int For Rev. 2008;10(4):573–90.
- 84. Owusu-Nimo F, Mantey J, Nyarko KB, Appiah-Effah E, Aubynn A. Spatial distribution patterns of illegal artisanal small scale gold mining (Galamsey) operations in Ghana: A focus on the Western Region. Heliyon. 2018;4(2):e00534. pmid:29511743
- 85. Mayer MJ, Gróf G. Extensive comparison of physical models for photovoltaic power forecasting. Appl Energy. 2021;283:116239.
- 86. Markovics D, Mayer MJ. Comparison of machine learning methods for photovoltaic power forecasting based on numerical weather prediction. Renew Sustain Energy Rev. 2022;161:112364.
- 87. Markowicz KM, Zawadzka-Manko O, Lisok J, Chilinski MT, Xian P. The impact of moderately absorbing aerosol on surface sensible, latent and net radiative fluxes during the summer of 2015 in Central Europe. J Aerosol Sci. 2021;151:105627.
- 88.
Murto S. The impact of aerosols on the sensible and latent heat fluxes in Beijing. Helsingfors Universitet; 2017.
- 89. Steiner AL, Mermelstein D, Cheng SJ, Twine TE, Oliphant A. Observed Impact of Atmospheric Aerosols on the Surface Energy Budget. Earth Interactions. 2013;17(14):1–22.
- 90.
Rosenfeld D. Aerosol-cloud interactions control of earth radiation and latent heat release budgets. Solar variability and planetary climates; 2007. p. 149–57.
- 91. Charlson RJ, Schwartz SE, Hales JM, Cess RD, Coakley JA Jr, Hansen JE, et al. Climate forcing by anthropogenic aerosols. Science. 1992;255(5043):423–30. pmid:17842894
- 92.
Rain D, Engstrom R, Ludlow C, Antos S. Accra Ghana: A city vulnerable to flooding and drought-induced migration; 2011. http://www.unhabitat.org/grhs/2011
- 93.
Ghana Statistical Service. 2019.
- 94. He J, Yu Y, Xie Y, Mao H, Wu L, Liu N, et al. Numerical model-based artificial neural network model and its application for quantifying impact factors of urban air quality. Water Air Soil Pollut. 2016;227:1–16.
- 95. Srivastava D, Singh RM. Groundwater system modeling for simultaneous identification of pollution sources and parameters with uncertainty characterization. Water Resour Manage. 2015;29(13):4607–27.
- 96. Yi Shang, Wah BW. Global optimization for neural network training. Computer. 1996;29(3):45–54.