Conformal PM2.5 Mapping Under Spatial Covariate Shift: Satellite-Reanalysis Fusion for Africa's Green Industrial Transition

Yaw Osei Adjei; Davis Opoku; Ephraim Abotsi; Kwadwo Owusu Amanqua; Oliver Kornyo; Elisha Soglo-Ahianyo; Cephas Anertey Abbey

cs.LGApr 5, 2026

Conformal PM2.5 Mapping Under Spatial Covariate Shift: Satellite-Reanalysis Fusion for Africa's Green Industrial Transition

Authors: Yaw Osei Adjei, Davis Opoku, Ephraim Abotsi, Kwadwo Owusu Amanqua, Oliver Kornyo, Elisha Soglo-Ahianyo, Cephas Anertey Abbey

Organizations: Kwame Nkrumah University of Science and Technology, Kumasi, Ghana · Dept. of Computer Science, Kwame Nkrumah Univ. of Science and Technology, Kumasi, Ghana

Abstract

Africa's green industrialization imperative demands reliable infrastructure for monitoring air quality. We present a satellite-reanalysis PM2.5 fusion system trained on 2,068,901 records from 404 monitoring locations in 29 African countries (OpenAQ, 2017-2022), combining LightGBM with leakage-resistant spatial cross-validation and conformal prediction to quantify predictions and their geographic applicability limits. Under 5-fold location-grouped spatial cross-validation, LightGBM achieves RMSE = 30.83 +/- 5.07 ug/m3, MAE = 14.54 +/- 1.66 ug/m3, R2 = 0.134 +/- 0.023, and macro F1 = 0.336 +/- 0.018. This R2 is substantially below random-split benchmarks (>0.90) but reflects true geographic generalisation difficulty rather than model failure. Split conformal prediction targeting 90% marginal coverage reveals severe East Africa degradation (actual PICP = 65.3% vs. nominal 90%), consistent with medium-strength covariate shift (humidity KS = 0.2237, sat_pblh KS = 0.2558). We operationalise these findings through regional reliability flags (High/Medium/Low/Unreliable) and a monitor prioritisation score directing infrastructure expansion toward highest-burden unmonitored populations, directly supporting Africa's green industrial transition and SDGs 3.9, 7.1.2, 9, 11.6.2, and 13.

Figures & tables

Panel of Figure 1 — Fig. 1: PM 2.5 monitoring station density (stations per million population, log scale). African sub-regions are 10–50 $\times$ more sparse than Europe and North America. Central and East Africa ( $<0.1$ stations/million) represent the most acute monitoring deserts. Dashed line indicates an aspirational minimum of 1 station per million. Sources: IQAir 2025 [ 2 ] ; Think Global Health 2024 [ 7 ] ; WHO v6.1 2024 [ 8 ] .

Panel of Figure 2 — Fig. 2: Data overview. (a) Top 10 African countries by monitoring location count (total 404 locations, 29 countries); (b) PM 2.5 records by year (total 2,068,901 records, 2017–2022, OpenAQ [ 33 ] ).

Group	Variables
Geographic	latitude, longitude
Temporal	month sin , month cos , harmattan_flag
Atmospheric	sat_aot, sat_no2, sat_pblh
Meteorological	temperature, humidity, pressure,
	wind_speed, precipitation, clouds
Demographic	population density

TABLE I: Feature Groups Used in the Final Model

Category	Range ( $\mu$ g/m 3 )	Health Implication
Good	[0, 12]	Negligible risk
Moderate	(12, 35]	Sensitive groups
Unhealthy for Sensitive (USG)	(35, 55]	Sensitive groups at risk
Unhealthy	(55, 150]	General population at risk
Very Unhealthy	(150, 250]	Widespread health risk
Hazardous	$(250,\infty)$	Emergency conditions

TABLE II: AQI-Style PM 2.5 Bins

Panel of Figure 5 — Fig. 3: (a) Predicted vs. observed PM 2.5 under 5-fold spatial CV; dashed line = perfect prediction, solid line = linear fit. (b) Normalised confusion matrix (row percentages) for six-bin AQI classification. Diagonal entries are in bold blue; off-diagonal spread shows classification difficulty particularly for Very Unhealthy and Hazardous classes.

Model	RMSE ( $\mu$ g/m 3 )	MAE ( $\mu$ g/m 3 )	R 2	Accuracy	Macro F1
Seasonal Naive	$33.24\pm 5.34$	$17.51\pm 1.09$	$-0.003\pm 0.007$	$0.397\pm 0.022$	$0.095\pm 0.004$
Ridge	$32.02\pm 5.10$	$16.20\pm 1.12$	$0.069\pm 0.009$	$0.415\pm 0.021$	$0.168\pm 0.013$
LightGBM	$30.83\pm 5.07$	$14.54\pm 1.66$	$0.134\pm 0.023$	$0.504\pm 0.031$	$0.336\pm 0.018$
XGBoost	$30.54\pm 5.63$	$14.43\pm 1.77$	$0.155\pm 0.025$	$0.487\pm 0.033$	$0.306\pm 0.018$

TABLE III: Model Performance: 5-Fold Location-Grouped Spatial CV (mean

\pm

std over folds)

Feature	KS Statistic	Severity
sat_pblh	0.2558	Medium
humidity	0.2237	Medium
sat_no2	0.0765	Low
sat_aot	0.0596	Low

TABLE IV: KS Covariate Shift Diagnostics: East Africa vs. Training

Panel of Figure 8 — Fig. 4: Regional conformal prediction performance (90 % nominal target, dashed line). North, West, and Southern Africa achieve near-nominal coverage; East Africa collapses to 65.3 % (24.7 pp shortfall), consistent with measured KS covariate shift (Table IV ). MPIW/2 (scaled for display) also peaks in East Africa (36.3 represents 72.6 $\mu$ g/m 3 full width), indicating wide but still insufficient intervals.

Sub-region	$n_{\text{loc}}$	R 2	RMSE	PICP	MPIW
			( $\mu$ g/m 3 )	(%)	( $\mu$ g/m 3 )
North Africa	48	$0.241$	$27.3$	$91.2$	$38.2$
West Africa	112	$0.187$	$29.1$	$90.5$	$41.5$
Central Africa	31	$0.062$	$32.7$	$82.1$	$58.3$
East Africa	97	$-0.083$	$33.9$	$65.3^{\dagger}$	$72.6$
Southern Africa	116	$0.218$	$28.4$	$89.8$	$42.1$

TABLE V: Regional Out-of-Fold Performance and Conformal Coverage (LightGBM; 90 % nominal target)

Panel of Figure 10 — Fig. 5: (a) Reliability flag distribution by African sub-region. East Africa shows the lowest reliability profile (12 % High, 24 % Unreliable). (b) Mean monitor prioritisation score (Eq. 2 ): East Africa ranks first, indicating high population exposure and high model uncertainty—the strongest case for new station deployment.

Panel of Figure 11 — Fig. 6: (a) Feature group ablation: RMSE increase when feature group is withheld from full LightGBM model. Meteorological features dominate; satellite AOT is second. (b) Regional SHAP instability for top-3 features. East Africa inversion (humidity $>$ sat_aot: 9.1 vs. 4.2) indicates attribution breakdown under covariate shift.

Panel of Figure 12 — Fig. 7: Global SHAP feature importance from out-of-fold LightGBM predictions ( $n=14{,}982$ , stratified by region and AQI bin). Satellite aerosol optical thickness (sat_aot) ranks first globally, consistent with its direct causal link to PM 2.5 loading. Meteorological covariates (humidity, temperature, pressure) collectively dominate—motivating their critical ablation impact (Fig. 6 a).

Panel of Figure 13 — Fig. 8: Green industrialisation policy cycle. Satellite–AI PM 2.5 fusion with conformal uncertainty quantification closes the measurement loop from industrial emissions through health impact assessment to policy-driven industrial greening—providing the evidence layer required for climate finance disbursement, just transition verification, and Agenda 2063 monitoring. SDG labels indicate where the system directly contributes to measurable global goals.

Indicator	Value	Source
Health cost of PM 2.5 as % of GDP (Africa avg.)	6.5 %	[ 10 ]
Projected health + productivity loss in 6 cities by 2040	US$138 B	[ 11 ]
EPA return per $1 on pollution control	30:1	[ 10 ]
Kenya GCF network investment	$9.3 M	[ 45 ]
Kenya: projected healthcare savings by 2040	$192 M (20:1)	[ 45 ]
SSA population exposed to unsafe PM 2.5	91.82 %	[ 32 ]

TABLE VI: Economic Case for AI-Based Air Quality Monitoring in Africa

Air Quality Prediction

Air quality prediction estimates future pollutant levels and maps local exposure to support health warnings, planning, and pollution control.

See all 42 papers

Cross-Resolution Attention Network for High-Resolution PM2.5 Prediction
Mar 12, 2026Ammar Kheder, Helmi Toropainen, Wenqing Peng +3Air Quality Prediction Spatiotemporal Forecasting
Department of Computational Engineering, LUT University, Finland · Atmospheric Modelling Centre Lahti (AMC-Lahti), Finland · Institute for Atmospheric and Earth System Research, University of Helsinki, P.O. Box 64, Helsinki 00014, Finland +1
Interpretable PM2.5 Forecasting for Urban Air Quality: A Comparative Study of Adaptive Time-Series Models
Mar 26, 2026Moazzam Umer Gondal, Hamad ul Qudous, Asma Ahmad Farhan +1Air Quality Prediction Time Series Forecasting
School of Computing, National University of Computer & Emerging Sciences (FAST), Lahore 54000, Pakistan · College of Computing and Informatics, Saudi Electronic University, Riyadh, 11673, Saudi Arabia
Global probabilistic forecasting of air pollution
Oct 19, 2026Ammar Kheder, Michael Boy, Zhi-song LiuAir Quality Prediction Probabilistic Forecasting
Department of Computational Engineering, LUT University, Lahti, Finland · Institute for Atmospheric and Earth System Research, University of Helsinki, Helsinki, Finland · Atmospheric Modelling Centre Lahti (AMC-Lahti), Lahti, Finland

Mar 12, 2026J/K move · Enter open · S save

Cross-Resolution Attention Network for High-Resolution PM2.5 Prediction

Ammar Kheder, Helmi Toropainen, Wenqing Peng, Samuel Antão+2

Department of Computational Engineering, LUT University, Finland · Atmospheric Modelling Centre Lahti (AMC-Lahti), Finland · Institute for Atmospheric and Earth System Research, University of Helsinki, P.O. Box 64, Helsinki 00014, Finland+1

Air Quality Prediction Spatiotemporal Forecasting Vision Transformer

Open paper AskPDF arXiv

Vision Transformers have achieved remarkable success in spatio-temporal prediction, but their scalability remains limited for ultra-high-resolution, continent-scale domains required in real-world environmental monitoring. A single European air-quality map at 1 km resolution comprises 29 million pixels, far beyond the limits of naive self-attention. We introduce CRAN-PM, a dual-branch Vision Transformer that leverages cross-resolution attention to efficiently fuse global meteorological data (25 km) with local high-resolution PM2.5 at the current time (1 km). Instead of including physically driven factors like temperature and topography as input, we further introduce elevation-aware self-attention and wind-guided cross-attention to force the network to learn physically consistent feature representations for PM2.5 forecasting. CRAN-PM is fully trainable and memory-efficient, generating the complete 29-million-pixel European map in 1.8 seconds on a single GPU. Evaluated on daily PM2.5 forecasting throughout Europe in 2022 (362 days, 2,971 European Environment Agency (EEA) stations), it reduces RMSE by 4.7% at T+1 and 10.7% at T+3 compared to the best single-scale baseline, while reducing bias in complex terrain by 36%.

Abstract

Air Quality Prediction

Explore similar work

Cross-Resolution Attention Network for High-Resolution PM2.5 Prediction

Interpretable PM2.5 Forecasting for Urban Air Quality: A Comparative Study of Adaptive Time-Series Models

Global probabilistic forecasting of air pollution