Ground-Level Near Real-Time Modeling for PM2.5 Pollution Prediction

Abstract: Air pollution is a worldwide public health threat that can cause or exacerbate many illnesses, including respiratory disease, cardiovascular disease, and some cancers. However, epidemiological studies and public health decision-making are stymied by the inability to assess pollution exposure impacts in near real time. To address this, developing accurate digital twins of environmental pollutants will enable timely data-driven analytics - a crucial step in modernizing health policy and decision-making. Although other models predict and analyze fine particulate matter exposure, they often rely on modeled input data sources and data streams that are not regularly updated. Another challenge stems from current models relying on predefined grids. In contrast, our deep-learning approach interpolates surface level PM2.5 concentrations between sparsely distributed US EPA monitoring stations in a grid-free manner. By incorporating additional, readily available datasets - including topographic, meteorological, and land-use data - we improve its ability to predict pollutant concentrations with high spatial and temporal resolution. This enables model querying at any spatial location for rapid predictions without computing over the entire grid. To ensure robustness, we randomize spatial sampling during training to enable our model to perform well in both dense and sparse monitored regions. This model is well suited for near real-time deployment because its lightweight architecture allows for fast updates in response to streaming data. Moreover, model flexibility and scalability allow it to be adapted to various geographical contexts and scales, making it a practical tool for delivering accurate and timely air quality assessments. Its capacity to rapidly evaluate multiple scenarios can be especially valuable for decision-making during public health crises.

• The paper proposes a ground-level, deep-learning model for predicting near real time PM 2.5 pollution utilizing data from sporadic environmental monitoring sites, without relying on slower, pre-defined gridded data.

Overview

This paper presents a grid-free, attention-based deep learning approach for interpolating surface-level PM2.5 concentrations across the contiguous United States (CONUS), built on the Senseiver architecture originally developed for sparse field reconstruction in fluid dynamics (2604.18973). The central motivation is near real-time exposure assessment: existing high-resolution PM2.5 products typically depend on modeled inputs (e.g., chemical transport model output, satellite retrievals) that are not updated promptly, and on predefined computational grids that impose resolution and cost trade-offs. The proposed model instead learns to interpolate directly between sparsely distributed EPA monitoring stations, incorporating readily available static and streaming covariates—meteorology, land cover, elevation, population density—and supports querying at arbitrary locations without recomputing an entire domain.

Methodology

The architecture adapts Senseiver's cross-attention mechanism between sensor observations and query points. Sensor inputs comprise lagged PM2.5 concentrations (a 15-day sliding window), meteorological variables from GridMET (4 km daily), land-cover embeddings learned from the 2013 NLCD (12-dimensional trainable embeddings), elevation from GMTED, LandScan day/night population, and Fourier-encoded geospatial coordinates. Query points carry only static variables and location encodings. Parameter recycling across lightweight attention blocks keeps the model small enough for rapid updates against streaming data.

Two design choices are notable. First, during training a target sensor is randomly selected as the query point, with nearby sensors dynamically sampled around it; this randomization forces the model to interpolate robustly across heterogeneous monitoring densities rather than overfitting to fixed spatial configurations. Second, uncertainty quantification is obtained at inference by Monte Carlo sampling of $M$ random sensor subsets $S^{(j)}$, computing predictions $$\hat{y}^{(j)} = f_\theta(x^* \mid \{(x_i,y_i)\}_{i\in S^{(j)}})$$, and aggregating into a predictive mean $\mu(x^*)$ and variance $\sigma^2(x^*)$ that capture epistemic sensitivity to sensor configuration without requiring a separate ensemble or Bayesian approximation.

All data streams were harmonized onto Uber H3 level-8 hexagons (~0.74 km²), with wind handled via U/V component decomposition to avoid circular-statistics artifacts. Lagged PM2.5 features were gap-filled using squared inverse-distance weighting (IDW2) over progressively expanding 5–50 km search radii, falling back to temporal linear interpolation. The dataset spans 2002–2020, drawn from 1,920 unique EPA AQS sites (1,893 after CONUS filtering), split 80/10/10 by spatiotemporal keys.

Results

Over the full 2002–2020 test set, the model achieves an MAE of 1.98 µg/m³ with $R^2 = 0.62$. Restricting to the dense prediction region of 0–20 µg/m³ yields $R^2 = 0.68$; excluding extreme values above 100 µg/m³ raises performance to $R^2 = 0.76$ with MAE of 1.7 µg/m³, indicating that the bulk of residual error is concentrated in rare high-concentration events. Year-over-year analysis shows MAE fluctuating between roughly 2.2 and 1.6 µg/m³ while MAPE increases steadily—a consequence of declining ambient PM2.5 levels rather than degrading absolute accuracy, though the authors note this trend is confounded by the growing number of observations per monitor (155,485 total observations in 2002 versus 256,319 in 2020).

Spatial cross-validation via leave-one-state-out (LOSO) on Utah—chosen for sparse monitoring, steep elevation gradients, and persistent winter inversions—shows only modest degradation: $R^2 = 0.670$ and MAE of 1.11 µg/m³ under LOSO versus $R^2 = 0.729$ and MAE of 1.01 µg/m³ when Utah is included in training. This is a meaningful generalization result for a deep learning interpolation model, though the authors concede that full leave-one-location-out validation across multiple states was not performed due to computational cost.

The uncertainty estimates show spatial concordance between predicted coefficient of variation and actual MAPE aggregated at H3 level-3 resolution, with a moderate Spearman correlation of 0.326 but weak linear fit ($S^{(j)}$0). The rank-based relationship is therefore informative for flagging low-reliability regions, but the CV should not be interpreted as a calibrated error magnitude.

The feature ablation study produces the paper's most counterintuitive finding: a minimal feature set (lagged PM2.5, land cover, time, location) performs comparably to the full multimodal set in aggregate and outperforms it in fall and winter in the western US. The authors attribute this to noise introduction and overfitting risk in data-sparse regions, concluding that parsimony can be preferable even when rich covariates are available. Both feature sets fail to capture spring trends adequately.

Independent validation uses two complementary strategies. First, the IMPROVE network—over 150 sites in remote Class I areas with minimal overlap with EPA regulatory monitors—serves as out-of-distribution ground truth for background locations. Second, a phenomenological case study of the Cameron Peak Fire (Colorado's largest recorded wildfire, >200,000 acres burned in 2020) shows the model reproducing coherent smoke plume evolution, topographic channeling into mountain valleys, and day-to-day variability consistent with satellite imagery from CIRA, with ground-truth sensors confirming localized accuracy during the extreme event.

Position relative to prior work

The approach departs from the dominant ensemble/data-fusion paradigm exemplified by Di et al.'s 1 km daily CONUS product, which integrates satellite AOD, meteorology, land use, and chemical transport model output. The authors deliberately exclude satellite AOD, citing its poor correlation with surface PM2.5 under complex vertical stratification, non-random missingness under cloud cover, and potential confounding in data-sparse regions. This is a defensible trade-off given the near real-time objective—AOD latency and gaps would compromise streaming deployment—but it means the model does not benefit from the broad spatial coverage AOD provides between monitors, which likely contributes to weaker western-US performance relative to AOD-fused alternatives such as Brokamp's and Di et al.'s models, whose behavior the results closely mirror in sparse regions.

Limitations and open questions

The paper is candid about several constraints. Performance degrades in the sparsely monitored western US, particularly in spring, where episodic events (wildfires, dust storms) dominate air quality; the LOSO experiment covers only one state, so geographic generalization claims rest on limited evidence. The uncertainty quantification is only moderately correlated with true error ($S^{(j)}$1 linearly), leaving calibration unresolved. Static covariates (2013 NLCD, 2016 LandScan) introduce temporal mismatch for early study years. The lagged-feature construction relies on IDW2 gap-filling, meaning the model partially trains on interpolated rather than measured values. Open questions include whether hybrid fusion with physical dispersion models or explicit temporal dynamics (recurrent or diffusion-based) would improve episodic-event fidelity, and whether the sensor-subset sampling distribution $S^{(j)}$2 can be optimized rather than treated as a hyperparameter.

Conclusion

This work demonstrates that a grid-free, sensor-adaptive attention model can deliver competitive PM2.5 interpolation accuracy (MAE ≈ 2 µg/m³, $S^{(j)}$3 up to 0.76 outside extremes) using only promptly available ground-based and contextual data streams, with built-in spatially resolved uncertainty and arbitrary-resolution querying suited to near real-time public health deployment. Its principal contribution is architectural flexibility rather than raw accuracy gains over established grid-based ensembles; the ablation finding that minimal features suffice in some regimes is a useful corrective to indiscriminate data fusion. Remaining weaknesses in sparse-monitoring regions and episodic events define the clearest targets for subsequent refinement.