Ghana is located in West Africa and lies between latitudes ∼4.5 and 11.5° N and longitudes 3.5° W and 1.5° E, encompassing an area of about 238 535 km2. The country is characterized by diverse geographic features, including coastal plains along the Gulf of Guinea, forested regions in the south, and savanna landscapes in the north. Major urban centers include Accra, Kumasi, Tamale, Sekondi-Takoradi, Cape Coast, Sunyani and the industrial port of Tema, representing important population and economic hubs with diverse emission characteristics.

Ghana has a typical tropical climate with distinct wet and dry seasons. Southern Ghana typically exhibits bimodal rainfall patterns due to its proximity to the Gulf of Guinea and the progression of the Intertropical Convergence Zone (ITCZ) (Nicholson, 2018), with major and minor wet seasons occurring between April–June and September–October, respectively. In contrast, northern Ghana has a unimodal wet season between April and October, peaking during August–September. Ghana experiences a meteorological phenomenon, known as the Harmattan, during mid-November to mid-March, which overlaps with the dry season. During this period, dry northeasterly winds transport large amounts of mineral dust from the Sahara Desert across West African countries, including Ghana (Alli et al., 2021). This results in extremely high concentrations of particulate matter across the country, leading to severely degraded air quality and reduced visibility, with impacts more pronounced in northern Ghana than in the south.

Ground-level measurements for PM_2.5 were collected from a network of 98 sensors spread across Ghana. The network consisted of 96 low-cost sensors, including 10 Airnote v2.0 (Blues Inc., Boston, MA), 21 Clarity Node-S Generation II (Clarity Movement Co., Berkeley, CA), 15 PurpleAir PA-II SD monitors (PurpleAir Inc., Draper, UT), 6 MODULAIR™ (QuantAQ Inc., Somerville, MA), and 44 MODULAIR™-PM monitors (QuantAQ Inc., Somerville, MA). PM_2.5 measurements from two BAMs (BAM-1020, Met One Instruments Inc., Grants Pass, Oregon) were also obtained, with one deployed in Techiman and another located at the U.S. Embassy in Accra, Ghana. The BAM-1020 uses a U.S. EPA-designated Federal Equivalent Method for measuring PM_2.5 and therefore was used as a reference monitor for calibrating raw PM_2.5 measurements from the low-cost sensors in the network. Table S1 in the Supplement details technical specifications of each monitor type, while Figs. S1 and S2 in the Supplement respectively illustrate the deployment locations and instrument-specific measurement periods for all sensors used in this study. PM_2.5 measurements collected from this sensor network between August 2018 and October 2024 were used as ground truth for model development. Since in situ measurements are available only from August 2018, PM_2.5 estimates for 2005–May 2018 were generated using models trained on the available surface observations and therefore cannot be directly validated against surface measurements from that period.

These sensors were strategically deployed across Ghana, with a significantly higher density in urban areas where population is largely concentrated, to capture high spatial variability in PM_2.5 concentrations arising from diverse sources, including vehicular emissions, open burning, cooking activities and industrial emissions (Kwarteng et al., 2020; Zhou et al., 2014). Air quality monitors were also deployed in suburban and rural regions of Ghana, where local air quality is strongly influenced by emissions from solid biomass fuel use for cooking, agricultural residue burning and dust from unpaved roads (Ofosu et al., 2013). This network architecture helps inform the model about PM_2.5 variability across northern, middle and southern belts of Ghana. The monitoring network had broader coverage across central and southern Ghana, with 35 sensors in the Kintampo area, 16 at Techiman, 20 at Accra, and 13 at Kumasi, while northern Ghana was represented by three sensors in Tamale (Table S2 and Fig. S1 in the Supplement). The sensors associated with Kintampo were distributed across the broader area surrounding Kintampo rather than being concentrated solely within the Kintampo town. Figure S1 depicts the locations of every monitor against a population count map retrieved from the LandScan project at Oak Ridge National Laboratory (Lebakula et al., 2025). Additional deployment details are available in Tables S2 and S3 in the Supplement.

Raw PM_2.5 measurements from each low-cost sensor, available at sub-hourly time resolution, were aggregated to hourly concentrations and subjected to custom-designed calibration models unique to each sensor type to derive quality assurance/quality control (QA/QC)-corrected exposure levels. The calibration procedures are further explained in Sect. S1 in the Supplement. These sensor-specific corrections were applied prior to model development to improve the consistency of PM_2.5 measurements across the different sensor types.

We developed a multi-model framework by integrating satellite retrievals, reanalysis products, and ground-based measurements to ensure complete coverage of PM_2.5 for the entire geographical region of Ghana and all days from 2005 to 2025 (Fig. 1). The framework consists of three complementary machine learning models with varying input features: the TROPOMI model, OMI model, and MERRA-2 model. All three models were trained using ground-level PM_2.5 measurements and incorporated meteorological variables from the ERA5-Land reanalysis dataset. The TROPOMI model derives PM_2.5 beginning 7 May 2018 by integrating AOD from MODIS and air quality parameters (UVAI, CO, HCHO, NO2, O3, and SO2) from the TROPOMI instrument. The OMI model also uses MODIS AOD but utilizes only UVAI, HCHO, NO2, O3, and SO2 retrievals from OMI to estimate PM_2.5 from 1 January 2005 to 6 May 2018, before TROPOMI measurements became available.

Since OMI and TROPOMI models rely heavily on satellite-derived observations, occasional cloud cover and instrument-specific limitations, including row anomalies in OMI (Torres et al., 2018) and retrieval uncertainties in TROPOMI (Veefkind et al., 2012), can result in missing or invalid satellite retrievals, preventing PM_2.5 estimates from both models for certain regions and time periods. The MERRA-2-based model was developed to mitigate these PM_2.5 gaps. It utilizes air quality features from MERRA-2 reanalysis dataset, including BCSMASS, OCSMASS, SO4SMASS, SO2SMASS, DUSMASS25, SSSMASS25, and TOTEXTTAU. Although MERRA-2 also contains meteorological variables, those features were instead obtained from the ERA5-Land reanalysis dataset due to availability of hourly variables at approximately 31 times finer grid sizes (0.1°×0.1°) which is more suitable for developing a high spatial resolution PM_2.5 model. The MERRA-2-based model therefore serves as a fallback framework for periods or locations where satellite retrievals from OMI or TROPOMI are unavailable, ensuring continuous spatiotemporal PM_2.5 coverage across Ghana from 2005–2025.

For each day and 1 km grid cell, the corresponding satellite-based model (TROPOMI or OMI) was used when the required satellite predictors were available. Grid cells where a satellite-based prediction could not be generated were then filled using the MERRA-2-based model. Thus, the final PM_2.5 product was constructed through a hierarchical prediction and gap-filling procedure. The OMI model was used from 1 January 2005 to 6 May 2018, and the TROPOMI model from 7 May 2018 onward, with the MERRA-2 model providing predictions only for remaining spatial or temporal gaps.

For each model type, geospatial machine learning algorithms, including LASSO, SVM, RF, XGBoost and MLP, were developed and evaluated to determine the optimal modeling approach for estimating daily PM_2.5 concentrations. Model performance was evaluated using cross-validation to ensure robust assessment and reduce the risk of overfitting. The probability distribution of PM_2.5 measurements from surface monitors exhibits a right-skewed distribution with a long tail towards higher concentrations (Fig. S6 in the Supplement). Therefore, the dataset was first divided into training (80 %) and testing (20 %) subsets using stratified sampling based on PM_2.5 concentration ranges. PM_2.5 values were grouped into discrete bins using thresholds of 35, 60, 100, 150, and 200 µgm-3, to ensure that low, moderate, and high pollution conditions were proportionally represented in both subsets. This stratification was aimed to improve model robustness by ensuring representation of episodic high-PM events, which is particularly important in geographical regions such as Ghana where the Harmattan season periodically leads to elevated particulate matter levels. Model hyperparameters were optimized using repeated five-fold cross-validation (5 folds repeated 5 times) applied to the training dataset, with the mean root mean square error (RMSE) across validation folds used as the loss function to identify the best-performing model configuration with minimal risk of overfitting. Model performance was assessed using r, R2, RMSE, MAE and mean bias. Mean-normalized RMSE (nRMSE) and MAE (nMAE) were computed by dividing the errors by the mean observed PM_2.5 concentrations, enabling comparison of model performance across environments with varying pollution levels. After determining the optimal hyperparameters, the model configuration was fixed and further evaluated using structured cross-validation to assess its ability to generalize across Ghana and time. Spatial and temporal cross-validation were used to reduce the influence of spatial and temporal autocorrelation and provide a more realistic estimate of model generalization.

Spatial cross-validation was conducted using a leave-one-location-out cross-validation approach, where data from one monitoring location were withheld as the test set while the model was trained using data from all remaining locations. This procedure was repeated until each location had served once as the test set. Locations with limited data coverage (Somanya and Sekondi) were excluded from the spatial cross-validation analysis because the sparse observations at these sites were insufficient for robust evaluation of model performance. This approach evaluates the model's ability to generalize to locations in Ghana where surface measurements are not available.

Temporal cross-validation was performed using the leave-one-year-out cross-validation approach, where observations from a single year were excluded from training and used exclusively for testing for each year from 2018 to 2024 of the surface measurement campaign. In each fold, the model was retrained using the fixed hyperparameters and evaluated on the held-out subset. This approach assesses the ability of the model to generalize across interannual variability in meteorology, emissions, and episodic pollution events, increasing confidence in the temporal robustness of PM_2.5 estimates for periods outside the training data. This step is particularly important because model training is based on observations from 2018–2024, while PM_2.5 concentrations are estimated retrospectively back to 2005, extending the application of the model beyond the period represented by the ground observations.

The optimized model from each predictor framework (TROPOMI, OMI, and MERRA) was subsequently applied to the complete gridded predictor dataset to generate spatially continuous daily PM_2.5 estimates across the spatiotemporal study domain. For each grid cell s and day t, predicted concentrations were obtained as PM2.5,s,t^=f(Xs,t), where Xs,t denotes the vector of satellite, meteorological, and cyclic predictors. The final output consisted of daily PM_2.5 maps at 0.01°×0.01° spatial resolution ranging from 2005 to 2025, enabling high-resolution spatiotemporal exposure assessment for subsequent analyses.

Daily PM_2.5 concentrations were estimated for 20 locations across Ghana by spatially aggregating selected 1 km gridded PM_2.5 estimates. The study locations included the regional capitals: Bolgatanga (Upper East), Nalerigu (North East), Wa (Upper West), Tamale (Northern), Damongo (Savannah), Dambai (Oti), Techiman (Bono East), Sunyani (Bono), Goaso (Ahafo), Kumasi (Ashanti), Ho (Volta), Sefwi Wiawso (Western North), Koforidua (Eastern), Cape Coast (Central), Sekondi-Takoradi (Western), and Accra (Greater Accra). Additional non-capital locations of interest consisted of Navrongo (Upper East), Kintampo (Bono East), Somanya (Eastern), and Tema (Greater Accra). For each location, a set of 1 km grid cells was selected based on the spatial extent and morphology of the surrounding built-up area, using Global Human Settlement Layer Urban Centre Database (GHS-UCDB) (Melchiorri et al., 2024) boundaries as the primary spatial reference where available and evaluating adjacent cells based on continuity of built-up development and surrounding spatial context (Sect. S2 and Fig. S7 in the Supplement). The selected cells were averaged to generate a representative daily PM_2.5 time series for each location.

To quantify the direction, magnitude, and statistical significance of long-term changes in surface PM_2.5 across Ghana, we applied the Seasonal Mann–Kendall (SMK) test and Sen's slope estimator to monthly mean PM_2.5 concentrations derived from the multi-model output for 2005–2025. Monthly means were calculated from daily 1 km gridded PM_2.5 concentrations, yielding 252 monthly observations per grid cell. The SMK test was used to account for the strong recurring seasonal cycle in PM_2.5 concentrations, particularly during the Harmattan season (Hirsch and Slack, 1984). Sen's slope estimator was used to quantify the magnitude of the monotonic trend in µgm-3yr-1 (Sen, 1968). For city-level trend analysis, the daily PM_2.5 concentrations aggregated over the selected 1 km grid cells for each location were first averaged to monthly concentrations. The SMK test and Seasonal Sen's slope estimator were then applied to the resulting monthly city-level time series to estimate the overall trend across the study period. The same approach was applied individually to the Harmattan (15 November–15 March) and non-Harmattan periods (16 March–14 November), with daily PM_2.5 concentrations first classified according to these date ranges and then aggregated to monthly mean concentrations.

Since the pixel-wise trend analysis involves spatially contiguous grid cells, we also evaluated spatial autocorrelation in the gridded PM_2.5 fields using Global Moran's I with a first-order rook neighborhood (Moran, 1950). Global Moran's I was calculated annually for the 0.01°×0.01° PM_2.5 fields for 2005–2025 and compared with that of the V6_GL_03 gridded PM_2.5 product from Shen et al. (2024) at 0.01 and 0.1° resolutions. Additional details are presented in Sect. S3 in the Supplement.

For the pixel-wise trend analysis, the Benjamini–Hochberg false discovery rate (FDR) procedure was applied to the SMK p-values across all valid grid cells. The resulting FDR-adjusted p-values (q-values) were used to identify statistically significant trends, with q<0.05 considered statistically significant (Benjamini and Hochberg, 1995). The SMK test and Sen's slope were applied independently to the monthly mean PM_2.5 time series at each 1 km grid cell across Ghana to produce spatially resolved maps of trend magnitude and significance. Additionally, spatially averaged time series were computed for the national domain and for distinct sub-regions of Ghana (northern, middle, and southern belts) to characterize large-scale changes in PM_2.5 burden over the 21-year study period. Further details of the pixel-wise trend analysis and FDR correction are provided in Sect. S4 in the Supplement.

The daily PM_2.5 concentrations collected by surface monitors during the measurement campaign (2018–2024) exhibited a mean concentration of 31.4 µgm-3, indicating consistently elevated pollution levels. The distribution of observed daily PM_2.5 concentrations is strongly right skewed, characterized by a long tail toward higher concentration values, as shown in Fig. S6. The 95th percentile PM_2.5 levels reached 80.0 µgm-3, 2.6 times the mean concentration which highlights the frequent occurrence of episodic high-pollution events. The large proportion of Harmattan-season observations at higher PM_2.5 concentrations suggests that these extreme values are strongly associated with Harmattan dust transport during the dry season. The skewed distribution indicates that while most observations fall within moderate concentration ranges, a smaller proportion of high-concentration events contributes disproportionately to the overall PM_2.5 distribution.

Figure S8 in the Supplement shows violin distributions of measured daily PM_2.5 concentrations across all cities with surface monitoring locations, revealing pronounced spatial heterogeneity in mean daily PM_2.5. The embedded box plots indicate that median concentrations at all locations range between 15–40 µgm-3, consistently exceeding the WHO 24 h guideline of 15 µgm-3 (World Health Organization, 2021).

Model performance was rigorously evaluated across five machine learning algorithms, LASSO, RF, SVM, XGBoost and MLP, for three model configurations TROPOMI model, OMI model, and MERRA model, under random, temporal and spatial cross-validation (CV) schemes. These algorithms represent different modeling approaches, including linear and kernel-based regression (LASSO and SVM), tree-based ensemble methods (RF and XGBoost), and neural networks (MLP). The MLP is a fully connected neural network, and its architecture used in this study is illustrated in Fig. S9 in the Supplement. Optimized hyperparameter information for each ML algorithm and model configuration is summarized in Tables S5–S7 in the Supplement. Model performance metrics, including Pearson correlation coefficient (r), coefficient of determination (R2), root-mean square error (RMSE) and mean-normalized RMSE (nRMSE), for all machine learning algorithms evaluated for the three model configurations are summarized in Table 1. Mean absolute error (MAE), and mean-normalized MAE (nMAE) are presented in Tables S8–S10 in the Supplement as additional error metrics.

Under random cross-validation, XGBoost achieved the strongest and most consistent performance across all three model configurations, with R2 of 0.72 for TROPOMI, OMI, and MERRA-2 models and RMSE values of 12.5, 12.8, and 12.6 µgm-3, respectively. Linear models (LASSO) consistently showed the weakest performance, with R2 approximately 0.09–0.12 lower than XGBoost and substantially higher RMSE across all cross-validation schemes, reflecting its limited ability to capture the non-linear relationships of surface PM_2.5 with meteorological and satellite-derived aerosol predictors in the heterogeneous emission environments in Ghana. Support vector machines (SVM) showed competitive random-CV performance but degraded substantially under spatial and temporal cross-validation. Among the non-linear approaches, tree-based ensemble methods RF and XGBoost performed comparably under random cross-validation, but XGBoost exhibited consistently lower bias (-0.01, 0.04, and -0.18 µgm-3 for TROPOMI, OMI and MERRA-2 models, respectively) and comparable or lower normalized errors across all CV schemes. The regression slopes in Fig. 2A–C (0.72–0.73) further indicate that XGBoost predictions closely track observed variability in PM_2.5 but tend to underestimate higher concentrations.

Consistent with the performances of spatiotemporal models in other studies, the models retained substantial predictive skill under temporal and spatial cross-validation despite a slight decrease in performance relative to random cross-validation (Chi et al., 2022). Under temporal cross-validation, XGBoost achieved R2 values of 0.57, 0.56, and 0.59 for the TROPOMI, OMI, and MERRA-2 models, with RMSE of 15.6, 15.6, and 15.2 µgm-3, respectively. The leave-one-year-out validation results show that the model maintained consistent predictive performance across different years, indicating that the learned predictor–PM_2.5 relationships are not strongly dependent on the meteorological and pollution characteristics of any single year. Under spatial cross-validation, which directly evaluates model performance at monitoring locations excluded from training, XGBoost R2 values were 0.57, 0.54, and 0.49 for TROPOMI, OMI, and MERRA models. The MLP model showed the largest degradation under spatial cross-validation, with R2 decreasing to 0.42, 0.46, and 0.44 for TROPOMI, OMI, and MERRA models, indicating poorer spatial generalization. In addition to its robust spatial and temporal generalization, XGBoost provides better interpretability through SHAP-based analysis, offering insights into the relative contributions of meteorological and aerosol predictors and further supporting its suitability for generating spatially continuous daily PM_2.5 maps across unmonitored regions and time periods in Ghana. Therefore, XGBoost was selected as the final model for further analysis. Figure 2A–C further illustrate the agreement between observed and predicted PM_2.5 concentrations for the best-performing model (XGBoost) under random CV, showing strong clustering along the 1:1 line and limited systematic bias across the concentration range.

Among the three model schemes, the MERRA model achieved the lowest temporal and spatial cross-validation RMSE values of 15.2 and 16.8 µgm-3, respectively. However, spatially cross-validated R2 for the MERRA model (0.49) is slightly lower than that of the other models (0.54–0.57), suggesting that the lower absolute RMSE is partly attributable to the reduced PM_2.5 dynamic range in this gap-filling dataset due to inherently larger grid sizes of MERRA-2 dataset. The higher temporally cross-validated R2 of 0.59 suggests that MERRA-2 provides useful information on diel variability of aerosol-related predictors through hourly estimates compared to only once-a-day measurements from TROPOMI and OMI. Nonetheless, the MERRA model provides PM_2.5 estimates during periods of satellite data unavailability, enabling complete spatiotemporal coverage.

To quantify the contribution of the MERRA-2 fallback to the final product, we assessed satellite data availability across the 1 km grid for each day from 2005–2025. During the OMI period (2005–6 May 2018), satellite predictors were available for an average of 58.3 % of grid cells, with the remaining 41.7 % requiring the MERRA-2 fallback. Following the introduction of TROPOMI on 7 May 2018, satellite availability increased substantially to 99.1 %, reducing the average MERRA-2 fallback to 0.9 %. Monthly satellite availability also exhibited seasonal variability, with greater reliance on the MERRA-2 fallback during periods of lower satellite coverage (Fig. S10 in the Supplement). Annual satellite availability and MERRA-2 fallback are summarized in Table S11 in the Supplement.

SHAP analysis plots in Fig. 2D–F provide insights into the influence of predictor variables on PM_2.5 concentrations. The results indicate that meteorological and seasonal predictors dominate PM_2.5 variability in Ghana. Relative humidity exhibited the strongest influence on model predictions, with low-humidity conditions strongly associated with elevated PM_2.5 concentrations. This asymmetric and wide range of SHAP responses under dry conditions suggests that low-RH acts as a strong indicator of Harmattan season and is associated with high ambient PM_2.5 concentrations, consistent with enhanced dust transport and low rainfall during the dry season. In contrast, high-humidity conditions during the wet seasons correspond predominantly to negative SHAP values, consistent with efficient wet scavenging and reduced aerosol loading under increased precipitation and enhanced aerosol removal.

The seasonal cyclic features represented by the sine and cosine transformations of DOY exhibited large absolute SHAP values distributed across both positive and negative directions, indicating that the model successfully captured the strong annual periodicity of PM_2.5 concentrations in Ghana. Wind speeds (U and V) and precipitation variables from ERA5-Land also showed important contributions, consistent with the roles of winds in dust transport and precipitation in wet deposition of aerosols. These results are broadly consistent with findings from a recent satellite-based PM_2.5 study in West Africa by Westervelt et al. (2025), which identified meteorological drivers and dust transport as major determinants of surface aerosol loading.

Satellite-derived aerosol indicators, particularly the ultraviolet aerosol index (UVAI), showed consistently positive SHAP contributions in both the TROPOMI and OMI models, indicating higher predicted PM_2.5 concentrations under elevated absorbing aerosol conditions. These conditions are commonly associated with mineral dust transport during the Harmattan season and episodic biomass burning, highlighting the importance of absorbing aerosol loading in shaping surface PM_2.5 variability. For the MERRA-2-based model, PM composition variables, particularly BC, OC, and dust, exhibited substantial importance alongside meteorological predictors. The strong contributions from dust-related variables are consistent with the influence of Harmattan-driven mineral dust transport, while BC and OC likely reflect contributions from combustion-related aerosol sources, including biomass burning and other anthropogenic emissions. SHAP values for AOD suggest that the column-integrated aerosol loading represented by AOD in the MERRA-2 reanalysis provides valuable predictive information, particularly during high-PM_2.5 episodes, similar to the role of UVAI in OMI- and TROPOMI-based models. Despite differences in predictor types among the three model configurations, the overall consistency in dominant drivers provides confidence that the modeling framework captures physically meaningful controls on PM_2.5 variability across seasons and regions.

The daily PM_2.5 concentrations derived at 1km×1km resolution for 2005–2025 were used for assessing spatial and long-term patterns. Daily PM_2.5 concentration distributions for Harmattan and non-Harmattan periods are illustrated in Fig. 3A. During the non-Harmattan season, the distribution is relatively narrow, with a mode around 20 µgm-3 and limited probability mass extending beyond 60 µgm-3, reflecting more consistent and moderate pollution conditions, which may be driven primarily by local and regional anthropogenic sources. In contrast, the Harmattan season distribution is characteristically broader, with a mode ∼69 µgm-3 and a pronounced long tail extending to around 150 µgm-3, strongly indicating episodic high-concentration events associated with large amounts of mineral dust transport from the Sahara and Sahel to the West African regions during this season. Notably, both distributions are almost entirely (>99 %) situated above the WHO 24 h air quality guideline of 15 µgm-3 (dashed red line), underscoring that PM_2.5 exceedances in Ghana are not confined to the dust season but represent a year-round public health concern due to high exposure to ambient particulate matter. The wider spread and heavier tail of the Harmattan-season distribution imply that PM_2.5 exposure among the Ghanaian population during this period is not only substantially more variable, but also consistently higher, with extreme pollution episodes frequently reaching concentrations 2–10 times higher than the WHO 24 h guideline.

These distributions for 20 locations, including regional capitals, in Ghana are shown in Fig. S11 in the Supplement. The northern region shows a broader PM_2.5 distribution during the Harmattan period compared with a narrower distribution and much lower mean in southern Ghana. Major cities of Ghana, such as Accra, Cape Coast, Sekondi–Takoradi and Kumasi, are mostly located in the South, and show a much narrower PM_2.5 distribution with less pronounced seasonal variability during this period. On the other hand, rural or suburban locations exhibit relatively broader PM_2.5 distribution, potentially reflecting greater variability in emission sources, including episodic agricultural burning and savanna fires during the dry season (Ofosu et al., 2013).

Fire activity is an additional important contributor to seasonal PM_2.5 variability in Ghana, particularly across northern and transitional ecological zones dominated by savanna land cover (Fig. S12 in the Supplement). The MODIS MCD12Q1 Version 6.1 Land Cover Type product (Friedl and Sulla-Menashe, 2022) for 2024 shows that grasslands (42 %), savannas (27 %), and woody savannas (10 %) together account for nearly 80 % of Ghana's land area, with these ecosystems concentrated primarily in northern Ghana and the forest–savanna transition belt (Fig. S13 in the Supplement). Fire detections from the MODIS MCD14ML active-fire product (Giglio et al., 2016) during 2005–2025, retaining observations with confidence ≥30%, show that fire occurrence is spatially concentrated in these savanna-dominated regions, consistent with widespread seasonal biomass burning associated with agricultural land management, pasture renewal, and accidental wildfires. Annual mean fire frequency and fire radiative power (FRP), a proxy for combustion intensity, both show pronounced seasonal dependence, with substantially higher activity and FRP during the dry Harmattan months relative to the wetter non-Harmattan months. The elevated frequency and intensity of fires during the Harmattan season likely enhance regional aerosol loading through increased emissions of carbonaceous particulate matter, contributing to the broader and more right-skewed PM_2.5 distribution observed during this period. Frequent fire occurrences, concentrated predominantly within the savanna regions of Ghana (Fig. S12), further suggest an important role of biomass burning in seasonal aerosol loading and PM_2.5 variability in Ghana, complementing the dominant influence of long-range Saharan dust transport during Harmattan conditions (Ofosu et al., 2013).

This north-south difference in PM_2.5 exposure levels is even more evident from monthly PM_2.5 climatology averaged over the 21-year study period (2005–2025), disaggregated by northern and southern regions of Ghana, as shown in Fig. 3B. Ghana is classified into northern and southern subregions, defined as regions above and below 8° N latitude, respectively (Klutse et al., 2020). A pronounced seasonal cycle can be observed for monthly mean PM_2.5 in Ghana, with peak concentrations occurring from November through March, coinciding with the Harmattan season. Typically, the North experienced 1.6–1.8 times higher PM_2.5 exposure than those in the South. PM_2.5 concentrations decline sharply with the onset of the major wet season in April–June, driven by enhanced wet removal of atmospheric particles and suppressed dust transport. A secondary minimum occurs during the minor wet season (September–October), separated from the major wet season by a brief mid-season dry period in July–August during which PM_2.5 concentrations exhibit a modest secondary peak. This bimodal wet season structure is more pronounced in southern Ghana, consistent with the bimodal rainfall regime associated with the seasonal northward and southward migration of the ITCZ over the coastal and forested zones of West Africa (Nicholson, 2018). In contrast, northern Ghana exhibits a unimodal rainfall pattern peaking in August and September (Fig. S14 in the Supplement), during which monthly mean PM_2.5 concentrations (19.9 µgm-3) are lower than those observed in the south (21.0 µgm-3).

Figure 4A shows the seasonal and annual evolution of PM_2.5 in Ghana over the 21-year period. Annual mean PM_2.5 has remained relatively stable over the study period with the Sen's slope estimate indicating an average annual increase of only 0.01 µgm-3 and seasonal slopes varying between 0.005–0.165 µgm-3yr-1. The Seasonal Mann–Kendall test suggests that these seasonal and yearly trends observed in Ghana are not statistically significant (p≥0.05). However, the spatial map of Sen's slope estimates at 1km×1km grid resolution in Fig. 4B indicates that a large proportion of Ghana exhibits statistically significant trends. Because the pixel-wise trends are spatially dependent, we evaluated spatial autocorrelation using Global Moran's I. The PM_2.5 fields exhibited strong positive spatial autocorrelation, with annual Global Moran's I values averaging approximately 0.995 (range: 0.992–0.997) across 2005–2025. Similar strong spatial autocorrelation was observed in the PM_2.5 product from Shen et al. (2024) at both 0.01 and 0.1° resolutions with Global Moran's I values of 0.999 and 0.989, respectively (Table S12 in the Supplement), indicating that strong spatial dependence is also characteristic of established gridded PM_2.5 products. To account for multiple comparisons across the spatial domain, the Benjamini–Hochberg FDR test was applied to the pixel-wise Mann–Kendall p-values. Following FDR correction, 66.93 % of valid grid cells exhibited statistically significant trends (q<0.05), compared with 70.14 % based on the uncorrected p<0.05 criterion (Table S13 in the Supplement). The Sen's slope estimates and p values calculated from the Seasonal Mann–Kendall analysis for the 20 study locations are summarized in Table S14 in the Supplement.

City scale trend plots in Fig. 5 show increasing PM_2.5 trends (0.05–0.11 µgm-3yr-1) over the 21-year study period in some suburban and rural regions such as Nalerigu, Wa, and Damongo in northern Ghana. Other regions exhibiting increasing PM_2.5 trends are Bono and Bono East, where cities such as Sunyani and Techiman may reflect increasing emissions associated with anthropogenic activities in these rapidly developing urban centers, in addition to contributions from Harmattan-related PM_2.5 episodes. In contrast, several southern cities exhibit statistically significant declining trends, with the strongest decreases observed in Somanya (-0.086 µgm-3yr-1), Ho (-0.072 µgm-3yr-1), and Koforidua (-0.069 µgm-3yr-1), followed by Cape Coast (-0.036 µgm-3yr-1), Sekondi-Takoradi (-0.030 µgm-3yr-1), and Tema (-0.023 µgm-3yr-1), which may be associated with improvements in vehicular fuel quality (Ayetor et al., 2025) and gradual adoption of cleaner household energy sources, potentially reflecting more effective policy implementation in the urbanized southern coastal regions of Ghana. Although these trends are statistically significant, their cumulative magnitudes over the 21-year period (approximately 0.5–1.7 µgm-3) are modest relative to the model's daily spatial prediction uncertainty, which ranges from approximately 7–21 µgm-3 RMSE across southern monitoring sites (Fig. S15 in the Supplement). Since the trend analysis is based on monthly rather than daily PM_2.5 concentrations, temporal averaging reduces the influence of unsystematic day-to-day prediction errors on the estimated long-term trends. To further assess temporal robustness, we evaluated model performance using leave-one-year-out cross-validation for 2018–2024, testing model generalization across interannual variability in meteorology, emissions, and episodic pollution events. A map of leave-one-location-out cross-validation RMSE at the monitoring locations is provided in Fig. S15 to contextualize model uncertainty in estimating daily PM_2.5 across Ghana. Changes in the characteristics of valid satellite retrievals or in their relationship with surface PM_2.5 over the longer record could introduce systematic differences in the retrospective estimates that cannot be directly evaluated in the absence of pre-2018 surface observations. Such differences could also influence the magnitude of estimated long-term trends and should therefore be considered when interpreting the retrospective record.

Overall, long-term changes in PM_2.5 across Ghana typically ranged between -0.16 and 0.14 µgm-3yr-1 (Fig. 4B). These trends suggest a persistent north–south gradient in PM_2.5 exposure over the 21-year period, with geographically divergent long-term trends superimposed on substantial interannual and multi-year variability. Positive trends in northern Ghana contrast with stable or declining trends in the south, suggesting regional differences in the factors influencing long-term PM_2.5 variability. This north–south pattern should be interpreted in the context of the spatial distribution of monitoring sites, particularly the lower monitoring density in northern Ghana. In northern Ghana, increases are consistent with the continued influence of Harmattan dust, regionally transported aerosols, and savanna fires, whereas several southern urban centers show stable or declining concentrations indicative of gradual changes in local emission patterns. At the national scale, the strong influence of interannual variability in Harmattan dust loading limits the detectability of statistically significant Ghana-wide trends, making it challenging to isolate long-term emission-driven changes from meteorological variability.

Trends during the non-Harmattan season are broadly consistent with the Seasonal Mann–Kendall results (Table S14) and provide additional insight into long-term PM_2.5 variability after accounting for the dominant seasonal dust cycle. Several southern cities, including Somanya, Ho, Koforidua, Tema, Cape Coast, and Sekondi-Takoradi, exhibit statistically significant declines in non-Harmattan PM_2.5 concentrations, indicating gradual declines in PM_2.5 concentrations outside the dust season. In contrast, northern cities such as Wa and Tamale show persistent positive non-Harmattan trends, indicating that factors beyond long-range dust transport, including local emissions and biomass burning, contribute to sustained PM_2.5 exposure levels throughout the year. The stronger declines observed in several southern urban centers may reflect, in part, the greater degree and spatial extent of urbanization in southern Ghana, where changes in urban-related emissions and household energy use may have a more pronounced influence on ambient PM_2.5. Additionally, although national initiatives have promoted the adoption of cleaner household fuels such as LPG, access to and adoption of LPG have historically been more limited in northern and rural areas, partly due to differences in fuel availability and distribution infrastructure (Dalaba et al., 2018). In northern Ghana, the influence of biomass burning and other regional sources may partially offset or obscure reductions in urban-related emissions. The spatial heterogeneity in non-Harmattan trends reinforces the north-south divergence observed in annual trends and highlights the combined influence of regional atmospheric processes and evolving local emission sources on long-term PM_2.5 variability across Ghana.

The spatial distribution of exceedance frequency relative to the Ghana EPA 24 h guideline of 35 µgm-3 (Ghana Standards Authority, 2019) is shown in Fig. 6. During 2005–2025, exceedances occur on approximately 15 %–20 % of days across southern Ghana and 35 %–42 % of days in northern regions (Fig. 6A). Restricting the analysis to non-Harmattan periods (Fig. 6B) substantially lowers exceedance frequency, with little to no exceedance in the south and fewer than 10 % of days exceeding the guideline in northern Ghana. In contrast, exceedance frequencies relative to the WHO 24 h guideline (15 µgm-3), shown in Fig. S16 in the Supplement, remain above 90 % across Ghana, including the southern coastal zone, reflecting the widespread and year-round burden of ambient PM_2.5 across Ghana in both Harmattan and non-Harmattan periods. Persistently high exceedance frequencies in major urban centers such as Accra, Kumasi, and Tema indicate that anthropogenic emissions contribute to sustained PM_2.5 exposure beyond the Harmattan season, highlighting the burden of ambient air pollution throughout the year across Ghana.

The city-level frequency distribution of daily PM_2.5 concentration ranges, shown in the bar plot in Fig. 6C, further quantifies the north-south contrast. The concentration ranges are defined using 24 h thresholds of the WHO Air Quality Guideline (15 µgm-3), WHO Interim Target 4 (25 µgm-3), Ghana EPA standard (35 µgm-3), WHO Interim Target 2 (50 µgm-3), and WHO Interim Target 1 (75 µgm-3) (Ghana Standards Authority, 2019; World Health Organization, 2021). During the Harmattan season, northern cities record substantially higher proportions of days exceeding 50 and 75 µgm-3 compared to southern cities, while even during non-Harmattan conditions, fewer than 10 % of days in any Ghanaian city fall below the WHO 24 h guideline of 15 µgm-3. The annual frequency corresponding to each concentration range for these cities is shown in Fig. S17 in the Supplement. Collectively, these exceedance patterns demonstrate that ambient PM_2.5 in Ghana constitutes a pervasive and chronic public health threat that cannot be attributed solely to episodic Harmattan dust events, underscoring the need for sustained emission reduction efforts across all seasons.

The 21-year mean spatial distribution of PM_2.5 across Ghana (Fig. 7) reveals a pronounced north–south gradient that is most extreme during the Harmattan season (Fig. 7B) and attenuates substantially during the wet seasons (Fig. 7C and  D). The 21-year annual mean (Fig. 7A) shows consistently elevated concentrations in the Upper East, Upper West, Northern, and Savannah regions compared to the forested south, reflecting the dominant influence of Harmattan-transported mineral dust from Saharan and Sahelian regions that disproportionately affects northern Ghana due to its open savanna landscape and exposure to dry northeasterly winds. During the major wet season (Fig. 7C), PM_2.5 concentrations drop significantly throughout the country, and the north-south gradient weakens considerably, driven by enhanced wet deposition and suppressed dust transport, while southern coastal cities retain discernible PM_2.5 elevations relative to the rural background, reflecting persistent contributions from traffic, open waste burning, and industrial activity. During the minor wet season (Fig. 7D), PM_2.5 remains relatively low but exhibits a more heterogeneous spatial pattern, with elevated concentrations extending across parts of central and southern Ghana. The city-level seasonal comparison (Fig. 7E) directly quantifies these contrasts: northern and central cities consistently experience substantially higher Harmattan-season PM_2.5 than their non-Harmattan levels, while southern coastal cities show comparatively muted seasonality, consistent with the dominance of year-round anthropogenic rather than dust-driven sources at those locations.

Decadal-scale changes in seasonal PM_2.5 concentrations between the two multi-year periods, 2005–2014 and 2015–2025, are shown in Fig. 8. During the Harmattan season, difference maps reveal a net increase in PM_2.5 across much of northern Ghana in the later period, consistent with the upward Sen's slope estimates for Wa, Tamale, Nalerigu, and Damongo (Table S14) and potentially reflecting intensified dust mobilization from Sahelian source regions under changing land use and climate conditions. In contrast, wet season PM_2.5 shows modest decreases in parts of southern Ghana between the two periods, particularly in and around major urban centers, corroborating the declining trends observed in Fig. 5 and Table S14 for cities such as Somanya, Ho, and Koforidua. In this inter-decadal comparison (Fig. 8), the most notable increases are concentrated along the northern belt, while decreases are evident in the southeastern coastal zone. Period-by-period comparisons at finer temporal resolution, comparing 2005–2009 with 2010–2014, 2010–2014 with 2015–2019, and 2015–2019 with 2020–2025, are provided in Figs. S18–S20 in the Supplement. Together, these analyses document persistent north–south differences in PM_2.5 exposure over the 21-year record, with contrasting long-term trends between a dust-burdened north and a south showing modest but measurable improvements, superimposed on substantial interannual and multi-year variability.