# Research Question 1
How can we decipher PM₂.₅ levels, independent of NO2 and SO2, using existing remote sensing, and satellite imagery?  

**Reviewers: Aakash Hariharan and Aishitha Pachipala**

## 1.1 Satellite-Based Estimation of Ground-Level PM₂.₅. 

### Key Findings

- {cite}`vanDonkelaar2006` presents one of the earliest physically based methods for estimating surface PM₂.₅ from satellite-derived AOD by explicitly linking column AOD to near-surface concentrations using a chemical transport model (GEOS-Chem).
- The method converts satellite-derived AOD from MODIS & MISR to PM2.5 using model-simulated aerosol vertical profiles, composition, and humidity effects, rather than relying on purely statistical AOD–PM regressions.

- {cite}`vanDonkelaar2010` develops an approach for combining MODIS and MISR AOD into a single improved estimate of AOD by exploiting the higher spatial coverage of MODIS and the superior aerosol-type and multi-angle retrieval capability of MISR, thereby reducing retrieval biases to provide a more robust AOD input for global surface PM₂.₅ estimation.

- {cite}`GuptaP2006` derives an empirical relationship between MODIS aerosol optical thickness (AOT/AOD) and surface PM₂.₅ mass concentration using collocated satellite and ground measurements across 26 sites in Sydney, Delhi, Hong Kong, New York City, and Switzerland. 
- The paper demonstartes that satellite derived AOD is a good surrogate for monitoring PM₂.₅ air quality, but also adds the importance of meteorological and other ancillary variables on PM2.5 estimation.

- {cite}`Dey2012` applies a satellite-based remote sensing framework to assess the spatial and temporal variability of surface PM₂.₅ over the Indian Subcontinent using MISR observations.
- Here AOD is converted to surface PM2.5 using spatially and seasonally varying conversion factors (η = PM₂.₅/AOD) derived from the GEOS-Chem chemical transport model, accounting for aerosol composition and vertical distribution.
- The study finds that MISR-derived PM₂.₅ shows a systematic low bias relative to ground observations, which is reduced through an empirical bias correction based on coincident measurements from Delhi and Kanpur, resulting in improved PM2.5 estimates across the region.

### Relevance to This Study

Collectively, these studies establish the methodological foundation for estimating ground-level PM₂.₅ from satellite-derived AOD using physically based approaches rather than purely empirical regressions. The work of van Donkelaar et al. (2006, 2010) demonstrates that incorporating aerosol vertical distribution, chemical composition, and humidity effects through chemical transport models such as GEOS-Chem is essential for accurately translating column AOD into surface PM₂.₅, particularly in regions with sparse ground monitoring. Gupta et al. (2006) further highlights that while satellite AOD is a useful proxy for PM₂.₅, its reliability depends strongly on meteorological conditions such as mixing height and relative humidity, underscoring the need to integrate ancillary atmospheric variables.

The application of this framework to the Indian Subcontinent by Dey et al. (2012) is especially relevant, as it confirms the transferability of MISR/MODIS-based AOD–PM2.5 conversion methods to South Asia, including Pakistan, while also emphasizing the importance of bias correction using limited ground observations. Together, these studies justify the use of satellite AOD combined with CTM-derived scaling factors and meteorological data to establish baseline PM2.5 distributions over Pakistan.

## 1.2 Hybrid Satellite–Model–Monitor PM₂.₅ Estimation at Global Scale.

### Key Findings

- {cite}`vanDonkelaar2016` develops a global PM2.5 estimation framework by combining satellite-derived AOD, chemical transport model (CTM) simulations (GEOS-Chem), and ground-based monitoring data using Bayesian geostatistical model.
- Satellite AOD is first converted to surface PM2.5 using CTM-based scaling factors that account for aerosol composition and vertical profiles, after which statistical calibration with ground monitors is applied to reduce regional biases.

## Relevance to This Study

This paper provides a well-established methodological foundation for converting satellite-derived AOD to PM2.5 using chemical transport models, while explicitly introducing a Bayesian geostatistical framework to integrate satellite estimates with ground-based observations. The Bayesian approach allows systematic quantification and correction of regional biases by formally combining multiple data sources and their uncertainties. This is particularly relevant for improving PM2.5 estimates in data-sparse regions such as Pakistan, where limited monitoring data can be optimally leveraged within a probabilistic framework.

## 1.3 Estimation of PM₂.₅ using Machine Learning Frameworks

### Key Findings 

- {cite}`DiQ2016` develops a high-resolution PM₂.₅ exposure model by integrating satellite AOD (MODIS), ground-based PM₂.₅ monitors, chemical transport model (CTM) outputs, meteorology, and land-use variables using a neural network.
- Results show that combining satellite data with CTMs and statistical learning can reliably capture fine-scale spatial and temporal variability in PM₂.₅, making the dataset suitable for large-scale health and exposure studies.

- {cite}`Wong2021` presents an integrated framework that combines land use regression with machine learning to estimate daily ground-level PM₂.₅ at high spatial and temporal resolution, using data from 73 air-quality monitoring stations in Taiwan (2006–2016) with external validation for 2017–2018.
- Among machine learning methods tested—Deep Neural Networks (DNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost); the Hybrid Kriging–LUR + XGBoost model performs best, achieving an adjusted R² of approximately 0.94. 
- Key predictors of PM₂.₅ include co-pollutants (SO₂, NO₂, O₃), meteorological factors, land-use characteristics (e.g., farmland, forest cover), proximity to emission-related infrastructure, and seasonal indicators, with forest cover exhibiting a negative association. Land use inventory data were not updated frequently, which may limit the model’s ability to capture temporal changes in land use and emission patterns over the study period. 

- {cite}`Ahmed2022` develops a deep learning based approach to estimate daily average ground level PM2.5 concentrations in Islamabad, Lahore, Karachi, and Peshawar of Pakistan using satellite imagery from the Sentinel 5P platform for for the period May 2019 to April 2020.
- A multi input convolutional neural network, referred to as P CNN, is developed to jointly process satellite derived images of aerosol index, methane, carbon monoxide, formaldehyde, nitrogen dioxide, ozone, and sulfur dioxide. 

- {cite}`RahmanM2022` develops a hybrid modeling framework to estimate source-specific PM₂.₅ concentrations at the census-tract level using chemical speciation data and machine-learning models.
- Absolute Principal Component Analysis (APCA) is applied to PM₂.₅ chemical composition data to identify major source factors, including motor vehicle emissions. The prediction framework integrates land-use variables, traffic-related variables, emissions inventories, meteorological variables, and satellite-derived PM2.5 constituents, with traffic characterized using road networks and Annual Average Daily Traffic (AADT).
- The paper estimates PM₂.₅ specifically attributed to motor vehicle traffic by separating total PM2.5 into source-specific components, which allows traffic-related pollution to be examined separately from other sources.

- {cite}`Chen2024` proposes a two stage PM₂.₅ estimation framework based on the Extreme Gradient Boosting algorithm to generate daily 1 km resolution PM₂.₅ maps across China without relying on satellite derived Aerosol Optical Depth.
- The two stage framework first introduces improved temporal encodings and terrain classification to represent spatiotemporal heterogeneity in PM₂.₅, and then incorporates an enhanced spatial autocorrelation term to mitigate the impact of uneven monitoring station distribution, resulting in strong performance under multiple cross validation schemes.
- Feature importance and SHAP analysis identify the improved spatial autocorrelation term as the most influential predictor, followed by temporal encodings and key meteorological variables such as temperature and total column ozone. 

- {cite}`Shen2026` develops a deep learning framework that integrates satellite AOD, GEOS-Chem chemical transport model outputs, land-use, and meteorological variables to estimate monthly total PM₂.₅ and its chemical components across North America at ~1 km resolution.
- Methodologically, the study employs a convolutional neural network (CNN) architecture to learn nonlinear relationships between predictors and surface PM₂.₅, while explicitly incorporating CTM-derived information as structured prior knowledge.
- Results show that incorporating geophysical a priori information from GEOS-Chem substantially improves model robustness in sparsely monitored regions and enhances uncertainty calibration.

## Relevance to This Study

These studies collectively demonstrate that machine learning frameworks provide a powerful extension to traditional satellite-based PM2.5 estimation methods by enabling the fusion of satellite observations, chemical transport model outputs, meteorological variables, land-use information, and ground-based measurements. The demonstrated ability of neural networks and tree-based models to downscale PM2.5 estimates supports the feasibility of generating high-resolution PM2.5 concentration surfaces in Pakistan, where monitoring coverage is limited and simple AOD–PM2.5 regressions are often insufficient.

Importantly, the reviewed frameworks show that PM2.5 can be estimated either in combination with CTM-informed satellite products or independently of AOD using spatiotemporal and contextual predictors, while remaining analytically separable from gaseous pollutants such as NO₂ and SO₂. This flexibility is directly relevant to the present study, as it allows PM2.5 to be modeled as a standalone exposure variable and subsequently analyzed in relation to transportation activity without conflating estimation methodology with pollutant attribution.

## Summary of Key Findings

The reviewed literature shows that ground-level PM2.5 can be reliably estimated from satellite observations by moving beyond simple empirical relationships and adopting physically informed and integrated modeling approaches. Studies using MODIS and MISR AOD demonstrate that chemical transport models such as GEOS-Chem are essential for translating column aerosol measurements into surface PM2.5 by accounting for aerosol vertical structure, composition, and meteorological influences. These approaches are particularly valuable in regions with limited ground monitoring, where satellite data provide the primary source of spatial coverage.

More recent work shows that combining satellite information with chemical transport models, ground observations, and machine learning techniques substantially improves the spatial and temporal resolution of PM₂.₅ estimates. Machine learning frameworks further enable PM2.5 to be modeled independently of gaseous pollutants such as NO2 and SO2, while allowing subsequent analysis of transportation-related influences. Together, these findings support the feasibility of estimating PM2.5 in Pakistan using remote sensing–based methods that are both physically grounded and adaptable to data-sparse environments.

## References
1. van Donkelaar, A., Martin, R. V., & Park, R. J. (2006). Estimating ground-level PM2.5 using aerosol optical depth determined from satellite remote sensing. Journal of Geophysical Research: Atmospheres, 111(D21), D21201. https://doi.org/10.1029/2005JD006996

2. van Donkelaar, A., Martin, R. V., Brauer, M., & Boys, B. L. (2010). Global estimates of ambient fine particulate matter concentrations from satellite-based aerosol optical depth: Development and application. *Environmental Health Perspectives, 118*(6), 847–855. https://doi.org/10.1289/ehp.0901623

3. van Donkelaar, A., Martin, R. V., Brauer, M., Hsu, N. C., Kahn, R. A., Levy, R. C., Lyapustin, A., Sayer, A. M., & Winker, D. M. (2016). Global estimates of fine particulate matter using a combined geophysical-statistical method with information from satellites, models, and monitors. Environmental Science & Technology, 50(7), 3762–3772. https://doi.org/10.1021/acs.est.5b05833

4. Di, Q., Kloog, I., Koutrakis, P., Lyapustin, A., Wang, Y., & Schwartz, J. (2016). Assessing PM2.5 exposures with high spatiotemporal resolution across the continental United States. *Environmental Science & Technology, 50*(9), 4712–4721. https://doi.org/10.1021/acs.est.5b06121

5. Dey, S., Di Girolamo, L., van Donkelaar, A., Tripathi, S. N., Gupta, T., & Mohan, M. (2012). Variability of outdoor fine particulate (PM2.5) concentration in the Indian Subcontinent: A remote sensing approach. Remote Sensing of Environment, 127, 153–161. https://doi.org/10.1016/j.rse.2012.08.021

6. Gupta, P., Christopher, S. A., Wang, J., Gehrig, R., Lee, Y. C., & Kumar, N. (2006). Satellite remote sensing of particulate matter and air quality assessment over global cities. Atmospheric Environment, 40(30), 5880–5892. https://doi.org/10.1016/j.atmosenv.2006.03.016

7. Wong, P.-Y., Lee, H.-Y., Chen, Y.-C., Zeng, Y.-T., Chern, Y.-R., Chen, N.-T., Lung, S.-C.-C., Su, H.-J., & Wu, C.-D. (2021). Using a land use regression model with machine learning to estimate ground-level PM2.5. Environmental Pollution, 277, 116846. https://doi.org/10.1016/j.envpol.2021.116846

8. Ahmed, M., Xiao, Z., & Shen, Y. (2022). Estimation of ground PM2.5 concentrations in Pakistan using convolutional neural network and multi-pollutant satellite images. Remote Sensing, 14(7), 1735. https://doi.org/10.3390/rs14071735

9. Rahman, M. M., & Thurston, G. D. (2022). A hybrid satellite and land use regression model of source-specific PM₂.₅ and PM₂.₅ constituents. Environment International, 163, 107233. https://doi.org/10.1016/j.envint.2022.107233

10. Chen, Q., Shao, K., & Zhang, S. (2024). Enhanced PM2.5 estimation across China: An AOD-independent two-stage approach incorporating improved spatiotemporal heterogeneity representations. Journal of Environmental Management, 368, 122107. https://doi.org/10.1016/j.jenvman.2024.122107

11. Shen, S., van Donkelaar, A., Jacobs, N., Li, C., & Martin, R. V. (2026). Enhancing estimation of fine particulate matter chemical composition across North America by including geophysical a priori information in deep learning with uncertainty quantification. ACS ES&T Air. https://doi.org/10.1021/acsestair.5c00251