CMAQ-OBS Dataset Overview

• CMAQ-OBS dataset is a fused air quality product that integrates CMAQ model simulations with ground observations to generate bias-corrected pollutant concentration fields.
• It employs advanced statistical methods, including spectral and spatiotemporal decomposition along with hierarchical calibration, to address observation-model discrepancies.
• The dataset supports epidemiological exposure assessment, real-time public health forecasting, and model evaluation across regions such as the US, East Asia, and Connecticut.

The CMAQ-OBS dataset is a fused air quality data product that integrates numerical model outputs from the Community Multiscale Air Quality (CMAQ) system with ground-based air quality observations, often complemented by local covariates, to produce high-resolution, spatiotemporally calibrated pollutant concentration fields. This class of datasets has been developed and applied in several major regions, including the contiguous United States for PM₂.₅ species, East Asia for real-time forecasting of PM and O₃, and Connecticut for fine-scale nitrogen dioxide (NO₂) exposure surfaces. The defining feature of CMAQ-OBS datasets is the explicit modeling and correction of the biases and scale discrepancies between CMAQ model simulations and observed measurements—enabling both accurate exposure assessment in epidemiology and operational public health forecasting.

1. Conceptual Foundation

CMAQ-OBS datasets are engineered by statistically fusing gridded CMAQ model outputs with ground-level observations to address the observation–model discordance arising from spatial aggregation, model error, and local unmodeled phenomena. The core methodological innovations include:

• Spectral or spatial decomposition of CMAQ fields to account for spatial scale–dependent bias and agreement structures (Guan et al., 2019).
• Hierarchical statistical modeling, often multivariate, that captures cross-dependencies among chemical species (e.g., PM₂.₅ constituents), allowing for cross-pollutant borrowing of information.
• Incorporation of site-level and environmental covariates for calibration, bias correction, and fine-scale downscaling (Gilani et al., 2019).
• Use of high-frequency or multi-resolution ground observations for both spatial and temporal refinement.
• The datasets serve dual purposes: as exposure surfaces for epidemiological studies and as validated training targets for advanced forecasting systems (Kang et al., 27 Nov 2025).

2. Data Sources and Spatial–Temporal Coverage

CMAQ-OBS construction draws from two primary data sources: (1) CMAQ model output, and (2) in situ measurement networks. Depending on the region and target pollutant, auxiliary covariates are also utilized. The following table provides comparative structural features of three referenced CMAQ-OBS products:

Region	Gridded Domain & Resolution	Temporal Resolution	Pollutants Included
Contiguous US	12 km × 12 km, 299×459 cells	Daily (2011 calendar yr)	PM₂.₅ + EC, OC, SO₄²⁻, NO₃⁻, NH₄⁺
East Asia	27 km regular grid over CN+KR	6-hourly (2016–2023)	PM₂.₅, PM₁₀, O₃
Connecticut (CT)	300 m × 300 m grid, CT only	Daily (1994–1995)	NO₂

CMAQ simulations are generated using meteorological drivers (e.g., WRF, MM5), region-specific emissions inventories, and chemical mechanisms (e.g., CB05, CB4). Site observation input includes regulatory monitoring networks (e.g., EPA AQS, AirKorea, China National), supplemented in some cases by specialized campaigns (e.g., Acid/Aerosol in CT).

3. Fusion and Downscaling Methodologies

Fusion approaches vary by region, variable, and research objective:

Multivariate Spectral Downscaling (Guan et al., 2019)

For contiguous US PM₂.₅ species, the approach involves:

• Log-transforming observations and CMAQ fields.
• Projecting CMAQ outputs into the Fourier domain, with filtering conducted via spline-basis functions to isolate scale-dependent features.
• A linear model regresses observed log-concentrations $Y_k(s)$ for species $k$ on filtered CMAQ features and a multivariate spatial residual $w_k(s)$; coefficients $\beta_{kjb}$ encode scale- and cross-species effects.
• The residual process $w(s)$ is modeled using a Linear Model of Coregionalization (LMC), parameterized as $w(s) = L v(s)$, with $L$ lower-triangular, and $v_i(s)$ independent GPs.
• Posterior inference is performed via MCMC, partitioned for computational scalability and aggregated with Consensus Monte Carlo.

Real-Time High-Resolution Gridding (Kang et al., 27 Nov 2025)

For East Asia, the fusion workflow is as follows:

• All CMAQ and OBS data are resampled to 6-hour intervals and interpolated to a common 27 km grid using nearest-neighbor or inverse-distance weighting (precise method not specified in the main text).
• Physically inconsistent values (e.g., negative concentrations) are removed using network QA flags.
• Output comprises gridded fields for PM₂.₅, PM₁₀, and O₃, temporally aligned and ready for both model initialization and as ground-truth for machine learning.

Spatiotemporal Calibration and Refinement (Gilani et al., 2019)

In Connecticut, the SCARR model fuses (i) spatially dense, temporally sparse passive sampler data and (ii) temporally dense, spatially sparse regulatory data with coarse CMAQ estimates:

• Step I: Spatial calibration using covariates (population, traffic, land use, seasonality) and CMAQ multiplicative bias, with negligible residual spatial error.
• Step II: State-space temporal calibration at monitor sites via a latent AR(1) process, fitted by maximum likelihood in R.
• The product is a gridded daily, calibrated NO₂ surface at 300 m resolution, decomposed into additive and multiplicative bias components, with full uncertainty propagation.

4. Output Formats, Validation, and Metadata

CMAQ-OBS datasets are output as time-stamped, gridded concentration fields in standard scientific formats:

• NetCDF is the de facto standard for 3D gridded data (dimensions: time, latitude, longitude; CF-compliant).
• Station-level predictions, bias fields, and diagnostic uncertainty estimates may be included as separate layers or files.
• Metadata includes variable names, species, units (e.g., μg m⁻³, ppb), grid reference, date/time, processing history, and version identification.

Validation is rigorous and quantitatively reported:

• Five-fold cross-validation at hold-out stations yields PM₂.₅ RMSE ≈ 2.56 μg m⁻³; corr ≈ 0.90 for CONUS (Guan et al., 2019).
• For East Asia, mean absolute error (CMAQ vs. OBS) is 21.33 μg m⁻³, a 59.5% bias reduction over CAMS (Kang et al., 27 Nov 2025).
• Monitor-site performance for CT NO₂ shows substantial MSE improvement at most sites via SCARR over raw CMAQ (Gilani et al., 2019).
• Model outputs are validated using error metrics such as RMSE, MAE, and MSE, with explicit formulas provided:

$$\operatorname{RMSE} = \sqrt{\frac{1}{N}\sum_{i=1}^{N}(y_{i} - \hat{y}_{i})^{2}}$$

$$\operatorname{MAE} = \frac{1}{N}\sum_{i=1}^{N} |y_{i} - \hat{y}_{i}|$$

• Diagnostic figures include spatial maps, seasonal averages, scale-dependent coherence curves, and animated time-series.

5. Applications and Impact

CMAQ-OBS datasets underpin a broad set of research and operational activities:

• Epidemiological Exposure Assessment: High-resolution, bias-corrected concentration fields enable large-scale studies of pollutant health impacts (Guan et al., 2019).
• Public Health Forecasting: In East Asia, the CMAQ-OBS product supports real-time, long-horizon (>48 h) air quality forecasts crucial for alerting vulnerable populations (Kang et al., 27 Nov 2025).
• Model Evaluation and Development: The fused data products provide accurate ground-truth for training and benchmarking machine-learning–based forecasting systems, outperforming standard global reanalyses (e.g., CAMS).
• Policy and Regulatory Support: The refined spatial and temporal granularity assists environmental agencies in air quality management, hotspot analysis, and policy effectiveness evaluation (Gilani et al., 2019).

6. Access, Reproducibility, and Future Directions

Data access, codebases, and reproducibility resources are region/product-dependent:

• CONUS PM₂.₅ species: R code for the spectral downscaling model and instructions are publicly available (https://github.com/yawenguan/multires); NetCDF files available on request (Guan et al., 2019).
• East Asia: Datasets can be accessed from https://github.com/KAIST-AI/FAKER-Air (dataset submodule “CMAQ-OBS”); inquires to [email protected] (Kang et al., 27 Nov 2025).
• Connecticut NO₂: Processed rasters and supplementary scripts (SAS, R) available from the authors; visual interactive access via Shiny application (Gilani et al., 2019).

A plausible implication is that future CMAQ-OBS datasets will increasingly leverage multi-source data fusion, deep learning–based assimilation, and higher-frequency observational networks to refine both spatial and temporal responsiveness. Emphasis on uncertainty quantification, metadata standardization, and user-driven customization (e.g., for chemical speciation or source attribution) is likely to accelerate.

7. Limitations and Challenges

Several methodological and operational challenges are inherent to the CMAQ-OBS approach:

• Spatial and temporal mismatch between observations and model output requires explicit statistical alignment strategies.
• Missing data in ground observations, uneven spatial coverage, and station-specific measurement artifacts necessitate robust, hierarchical modeling to avoid ad hoc imputation.
• While spectral and spatiotemporal calibration strategies yield substantial improvements, model transferability across regions and climates remains an ongoing research area.
• Data access, particularly for high-resolution ground-truth and emissions inventories, may be restricted by national or regional policies.

In summary, the CMAQ-OBS dataset paradigm provides rigorously validated, bias-corrected, spatiotemporally explicit air quality fields by fusing state-of-the-art chemical transport modeling with heterogenous ground observations—supporting critical advances in environmental health science, air quality forecasting, and regulatory science worldwide (Guan et al., 2019, Kang et al., 27 Nov 2025, Gilani et al., 2019).