arxiv: 2604.18973 · v1 · submitted 2026-04-21 · 📊 stat.AP · cs.LG

Recognition: unknown

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

Zachary R. Fox , Janet O. Agbaje , Dakotah Maguire , Javier E. Santos , Jeremy Logan , Maggie Davis , Rima Habre , Jim VanDerslice

show 1 more author

Heidi A. Hanson

Authors on Pith no claims yet

Pith reviewed 2026-05-10 02:02 UTC · model grok-4.3

classification 📊 stat.AP cs.LG

keywords PM2.5deep learningair pollutionspatial interpolationgrid-freenear real-timeenvironmental modeling

0 comments

The pith

A deep-learning model predicts surface-level PM2.5 concentrations at any location by interpolating sparse monitoring data without a fixed grid.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a deep learning approach for predicting ground-level PM2.5 pollution concentrations in near real time across the US. Unlike traditional methods that use fixed grids or rely on infrequently updated modeled data, this model interpolates between sparse EPA monitoring stations using additional topographic, meteorological, and land-use datasets. By randomizing the spatial locations sampled during training, the model maintains performance in both well-monitored and under-monitored areas, and its compact structure supports quick updates with incoming data streams. Successful implementation would allow on-demand pollution estimates at any point for timely public health analysis and policy decisions.

Core claim

The central claim is that a deep-learning model can interpolate surface level PM2.5 concentrations between sparsely distributed US EPA monitoring stations in a grid-free manner by incorporating topographic, meteorological, and land-use data, thereby enabling high spatial and temporal resolution predictions queryable at any location. Random spatial sampling during training ensures robustness across dense and sparse regions, while the lightweight architecture facilitates fast updates with streaming data for near real-time deployment and adaptability to various scales.

What carries the argument

Grid-free deep neural network interpolation of PM2.5 using auxiliary topographic, meteorological, and land-use data with randomized spatial sampling for training.

If this is right

Predictions can be generated instantly at any specific geographic location without processing an entire grid.
The model supports near real-time updates as new monitoring data becomes available due to its lightweight design.
It can be adapted to different geographical areas and scales where similar data sources exist.
Rapid evaluation of multiple pollution scenarios aids decision-making in public health emergencies.
Higher spatial and temporal resolution is achieved compared to grid-dependent models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same architecture might apply to other pollutants or environmental variables with analogous data inputs.
Combining this with satellite or mobile sensor data could enhance coverage in remote or rapidly changing areas.
Direct linkage to health outcome datasets could enable real-time exposure risk mapping for vulnerable populations.
Validation against historical pollution events would test its effectiveness for crisis response.

Load-bearing premise

The model assumes that randomizing spatial sampling during training will enable accurate performance in regions with both high and low densities of monitoring stations, and that its simple structure will allow accurate fast updates from live data streams.

What would settle it

If tests reveal significantly higher prediction errors in sparsely monitored rural areas compared to dense urban zones, or if incorporating new streaming data increases errors rather than maintaining accuracy, the central claims would not hold.

read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This is a practical engineering paper on grid-free PM2.5 interpolation via deep learning with auxiliary features; it has some held-out validation but thin baseline comparisons.

read the letter

The main thing here is a coordinate-based neural net that takes EPA station locations plus topographic, meteorological, and land-use inputs to output pointwise PM2.5 estimates without a fixed grid. Training randomizes which stations are seen in each batch so the model stays accurate in both dense and sparse areas, and the architecture is kept small enough for incremental updates on streaming data. The full text supplies the loss formulation, implementation choices, and quantitative results on held-out stations, which moves it past the abstract-only limitations noted in the initial read. That setup directly tackles the need for fast, location-specific exposure estimates in public health work. The randomization step and auxiliary data integration are straightforward ways to improve coverage where monitors are uneven, and the grid-free querying is a genuine operational advantage over raster-based alternatives. The paper does a decent job describing how the model could adapt to other regions or scales. The main soft spot is the evaluation. Results on held-out stations are reported, but direct comparisons to established spatial methods like kriging or inverse-distance weighting are limited or absent, so the incremental gain from the deep model is not fully quantified. Timing numbers for streaming updates are also missing, which leaves the near-real-time claim plausible but unproven in practice. No circularity or hidden contradictions show up in the argument. This is aimed at environmental modelers and exposure scientists who need workable tools rather than theorists. It is solid enough on methods and data use to deserve referee time, even if the authors will likely need to add stronger baselines and runtime tests.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a coordinate-based deep learning model to interpolate surface-level PM2.5 concentrations from sparse US EPA monitoring stations in a grid-free manner. Auxiliary topographic, meteorological, and land-use datasets are incorporated to achieve high spatial and temporal resolution, with spatial sampling randomization during training to ensure performance across dense and sparse monitoring regimes. The lightweight architecture is presented as enabling fast incremental updates with streaming data and adaptability to different geographical contexts.

Significance. If the held-out station results and implementation details hold under scrutiny, the work offers a practical, queryable alternative to grid-based models for near real-time air quality mapping. The use of readily available auxiliary data and emphasis on robustness to varying station density could support improved exposure assessment in epidemiological studies and public health decision-making during crises.

major comments (2)

[§4, Table 2] §4 (Results), Table 2: The quantitative metrics on held-out stations (e.g., RMSE and correlation) are reported without direct comparisons to standard baselines such as inverse-distance weighting, kriging, or other coordinate-based ML models; this weakens the claim that the auxiliary data and randomization yield meaningful improvements.
[§3.2] §3.2 (Model and Loss): The loss formulation and training procedure do not explicitly detail how the spatial sampling randomization is implemented (e.g., as a data augmentation step or modified objective), which is load-bearing for the robustness claim across density regimes.

minor comments (2)

[Abstract] The abstract states performance benefits but omits any numerical results; moving a concise summary of key metrics (e.g., from Table 2) into the abstract would improve accessibility.
[§3.1] Notation for the coordinate inputs and auxiliary feature vectors is introduced without a consolidated table; a single notation summary would aid readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive summary and recommendation for minor revision. The comments highlight opportunities to strengthen the manuscript, and we address each point below with plans for revision.

read point-by-point responses

Referee: [§4, Table 2] §4 (Results), Table 2: The quantitative metrics on held-out stations (e.g., RMSE and correlation) are reported without direct comparisons to standard baselines such as inverse-distance weighting, kriging, or other coordinate-based ML models; this weakens the claim that the auxiliary data and randomization yield meaningful improvements.

Authors: We agree that direct baseline comparisons are needed to substantiate the improvements from auxiliary inputs and randomization. In the revised manuscript, we will expand Table 2 and Section 4 to include results from inverse-distance weighting, ordinary kriging, and at least one other coordinate-based model (e.g., a simple MLP or Gaussian process regressor) trained on the same held-out stations and auxiliary features. This will allow quantitative assessment of gains in RMSE, correlation, and robustness across density regimes. revision: yes
Referee: [§3.2] §3.2 (Model and Loss): The loss formulation and training procedure do not explicitly detail how the spatial sampling randomization is implemented (e.g., as a data augmentation step or modified objective), which is load-bearing for the robustness claim across density regimes.

Authors: We acknowledge that the current description in §3.2 is insufficiently explicit. In the revision, we will clarify that spatial sampling randomization is implemented as a data-augmentation step: for each training batch, a random subset of stations (with size drawn from a distribution reflecting dense-to-sparse regimes) is selected, the model is queried only at those locations, and the loss (MSE on observed PM2.5) is computed solely on the sampled points. This procedure is repeated across epochs to encourage generalization; we will also provide pseudocode and hyperparameter ranges for the sampling distribution. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper presents a deep-learning interpolation model for PM2.5 that ingests station locations plus auxiliary topographic, meteorological, and land-use features. Training uses randomized spatial sampling on external EPA and auxiliary datasets, with a lightweight architecture for streaming updates. No equations, self-definitional reductions, fitted-input predictions, or load-bearing self-citations appear in the provided text; the central claims rest on standard supervised learning from independent data sources and are therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities are stated. The central claim implicitly rests on the untested assumption that randomized spatial sampling produces robust generalization across monitoring densities.

axioms (1)

domain assumption Randomized spatial sampling during training produces a model that generalizes to both dense and sparse real-world monitoring configurations
Stated in the abstract as the mechanism for robustness in varied regions.

pith-pipeline@v0.9.0 · 5606 in / 1357 out tokens · 57691 ms · 2026-05-10T02:02:10.248259+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 4 canonical work pages

[1]

Journal of Internal Medicine291(6), 779– 800 (2022)

Bont, J., Jaganathan, S., Dahlquist, M., Persson, ˚A., Stafoggia, M., Ljungman, P.: Ambient air pollution and cardiovascular diseases: An umbrella review of systematic reviews and meta-analyses. Journal of Internal Medicine291(6), 779– 800 (2022)

2022
[2]

Frontiers in Endocrinology9, 680 (2018)

Hamanaka, R.B., Mutlu, G.M.: Particulate matter air pollution: effects on the cardiovascular system. Frontiers in Endocrinology9, 680 (2018)

2018
[3]

American Journal of Public Health108(S2), 123–130 (2018)

Requia, W.J., Adams, M.D., Arain, A., Papatheodorou, S., Koutrakis, P., Mah- moud, M.: Global association of air pollution and cardiorespiratory diseases: a systematic review, meta-analysis, and investigation of modifier variables. American Journal of Public Health108(S2), 123–130 (2018)

2018
[4]

Environmental Science & Technology Letters11(11), 1220–1226 (2024)

Wang, Y., Marshall, J.D., Apte, J.S.: US ambient air monitoring network has inadequate coverage under new PM2.5 standard. Environmental Science & Technology Letters11(11), 1220–1226 (2024)

2024
[5]

EBioMedicine93(2023)

Vilcassim, R., Thurston, G.D.: Gaps and future directions in research on health effects of air pollution. EBioMedicine93(2023)

2023
[6]

Environmental Protection Agency: National ambient air quality standards (NAAQS)

U.S. Environmental Protection Agency: National ambient air quality standards (NAAQS). Technical report, U.S. Environmental Protection Agency (2009)

2009
[7]

Atmospheric Measurement Techniques14(6), 4617–4637 (2021)

Barkjohn, K.K., Gantt, B., Clements, A.L.: Development and application of a United States-wide correction for PM2.5 data collected with the PurpleAir sensor. Atmospheric Measurement Techniques14(6), 4617–4637 (2021)

2021
[8]

Atmospheric Measurement Techniques Discussions2019, 1–33 (2019)

Ardon-Dryer, K., Dryer, Y., Williams, J.N., Moghimi, N.: Measurements of PM2.5 with PurpleAir under atmospheric conditions. Atmospheric Measurement Techniques Discussions2019, 1–33 (2019)

2019
[9]

Environmental Protection Agency: Air Quality System Data Mart

U.S. Environmental Protection Agency: Air Quality System Data Mart. Avail- able at https://www.epa.gov/outdoor-air-quality-data. Accessed April 25, 2024 (2024)

2024
[10]

Environmental Protection Agency: How the NowCast Works: The Now- Cast method reports the Air Quality Index (AQI) in real time for particu- late matter (PM)

U.S. Environmental Protection Agency: How the NowCast Works: The Now- Cast method reports the Air Quality Index (AQI) in real time for particu- late matter (PM). Available at https://www.epa.gov/sites/default/files/2018-01/ 20 documents/nowcastfactsheet.pdf (2014)

2018
[11]

Environmental Protection Agency: AirNow Knowledge Base: How are the map contours made? What interpolation method is used? Available at https: //usepa.servicenowservices.com/airnow

U.S. Environmental Protection Agency: AirNow Knowledge Base: How are the map contours made? What interpolation method is used? Available at https: //usepa.servicenowservices.com/airnow. Accessed May 4, 2024 (2024)

2024
[12]

Nature Machine Intelligence5(11), 1317–1325 (2023)

Santos, J.E., Fox, Z.R., Mohan, A., O’Malley, D., Viswanathan, H., Lubbers, N.: Development of the Senseiver for efficient field reconstruction from sparse observations. Nature Machine Intelligence5(11), 1317–1325 (2023)

2023
[13]

American Review of Respiratory Disease 145(3), 600–604 (1992) https://doi.org/10.1164/ajrccm/145.3.600

Schwartz, J., Dockery, D.W.: Increased mortality in philadelphia associated with daily air pollution concentrations. American Review of Respiratory Disease 145(3), 600–604 (1992) https://doi.org/10.1164/ajrccm/145.3.600

work page doi:10.1164/ajrccm/145.3.600 1992
[14]

Dockery, D.W., Pope, C.A., Xu, X., Spengler, J.D., Ware, J.H., Fay, M.E., Ferris, B.G.J., Speizer, F.E.: An association between air pollution and mortality in six u.s. cities. New England Journal of Medicine329(24), 1753–1759 (1993) https: //doi.org/10.1056/NEJM199312093292401

work page doi:10.1056/nejm199312093292401 1993
[15]

American Journal of Respiratory and Critical Care Medicine150(5), 1234–1242 (1994) https://doi.org/10.1164/ ajrccm.150.5.7952546

Schwartz, J., Dockery, D.W., Neas, L.M., Wypij, D., Ware, J.H., Spengler, J.D., Koutrakis, P., Speizer, F.E., Ferris, B.G.J.: Acute effects of summer air pollution on respiratory symptom reporting in children. American Journal of Respiratory and Critical Care Medicine150(5), 1234–1242 (1994) https://doi.org/10.1164/ ajrccm.150.5.7952546

1994
[16]

Environmental Science & Technology47(13), 7233–7241 (2013)

Beckerman, B.S., Jerrett, M., Serre, M., Martin, R.V., Lee, S.-J., Van Donke- laar, A., Ross, Z., Su, J., Burnett, R.T.: A hybrid approach to estimating national scale spatiotemporal variability of PM2.5 in the contiguous United States. Environmental Science & Technology47(13), 7233–7241 (2013)

2013
[17]

Science of the Total Environment897, 166178 (2023)

Panneerselvam, B., Ravichandran, N., Dumka, U.C., Thomas, M., Charoenlerk- thawin, W., Bidorn, B.: A novel approach for the prediction and analysis of daily concentrations of particulate matter using machine learning. Science of the Total Environment897, 166178 (2023)

2023
[18]

Applied Intelligence 53(15), 18319–18332 (2023) https://doi.org/10.1007/s10489-022-04418-y

Chen, J., Yuan, C., Dong, S., Feng, J., Wang, H.: A novel spatiotemporal multi- graph convolutional network for air pollution prediction. Applied Intelligence 53(15), 18319–18332 (2023) https://doi.org/10.1007/s10489-022-04418-y

work page doi:10.1007/s10489-022-04418-y 2023
[19]

Scientific Reports10(1), 20988 (2020)

Xiao, F., Yang, M., Fan, H., Fan, G., Al-Qaness, M.A.: An improved deep learning model for predicting daily PM2.5 concentration. Scientific Reports10(1), 20988 (2020)

2020
[20]

Envi- ronmental Science & Technology52(22), 13260–13269 (2018) https://doi.org/10

Xiao, Q., Chang, H.H., Geng, G., Liu, Y.: An ensemble machine-learning model to predict historical PM2.5 concentrations in China from satellite data. Envi- ronmental Science & Technology52(22), 13260–13269 (2018) https://doi.org/10. 1021/acs.est.8b02917 21

2018
[21]

Environment International130, 104909 (2019)

Di, Q., Amini, H., Shi, L., Kloog, I., Silvern, R., Kelly, J., Sabath, M.B., Choirat, C., Koutrakis, P., Lyapustin, A.,et al.: An ensemble-based model of PM2.5 concentration across the contiguous United States with high spatiotemporal resolution. Environment International130, 104909 (2019)

2019
[22]

Science of the Total Environment751, 141813 (2021)

Zhang, H., Zhan, Y., Li, J., Chao, C.-Y., Liu, Q., Wang, C., Jia, S., Ma, L., Biswas, P.: Using Kriging incorporated with wind direction to investigate ground-level PM2.5 concentration. Science of the Total Environment751, 141813 (2021)

2021
[23]

Environmental Advances7, 100155 (2022)

Brokamp, C.: A high resolution spatiotemporal fine particulate matter exposure assessment model for the contiguous united states. Environmental Advances7, 100155 (2022)

2022
[24]

Technical report, Utah State University Extension (2012)

Gillies, R.R., Ramsey, R.D.: Climate of Utah. Technical report, Utah State University Extension (2012). Available at https://extension.usu.edu/rangelands/ files/RRU Section Five.pdf

2012
[25]

Geological Survey: Geographic information systems (GIS) data for the national atlas of the United States

U.S. Geological Survey: Geographic information systems (GIS) data for the national atlas of the United States. Open-File Report 2011-1073, United States Geological Survey (2011). Available at https://pubs.usgs.gov/publication/ ofr20111073

2011
[26]

University of California Merced

Abatzoglou, J.T.: GridMET: Gridded Surface Meteorological Dataset. University of California Merced. http://www.climatologylab.org/gridmet.html (2023)

2023
[27]

GitHub repository

Uber Technologies, Inc.: H3: A hexagonal hierarchical geospatial indexing system. GitHub repository. Version 4.1.0. https://github.com/uber/h3 (2023)

2023
[28]

Geological Survey: NLCD Land Cover Classification Legend

U.S. Geological Survey: NLCD Land Cover Classification Legend. Available at https://www.usgs.gov/media/images/nlcd-land-cover-classification-legend (2024)

2024
[29]

Oak Ridge National Laboratory

Rose, A., Weber, E., Moehl, J., Laverdiere, M., Yang, H., Whitehead, M., Sims, K., Trombley, N., Bhaduri, B.: LandScan USA 2016. Oak Ridge National Laboratory. Data set. Available at https://doi.org/10.48690/1523377 (2017)

work page doi:10.48690/1523377 2016
[30]

Atmospheric Chemistry and Physics21(22), 16775–16791 (2021)

Zhai, S., Jacob, D.J., Brewer, J.F., Li, K., Moch, J.M., Kim, J., Lee, S., Lim, H., Lee, H.C., Kuk, S.K.,et al.: Relating geostationary satellite measure- ments of aerosol optical depth (AOD) over East Asia to fine particulate matter (PM2.5): insights from the KORUS-AQ aircraft campaign and GEOS-Chem model simulations. Atmospheric Chemistry and Physics21(...

2021
[31]

Environmental Protection Agency: Prevention of Significant Deteri- oration (PSD) Basic Information

U.S. Environmental Protection Agency: Prevention of Significant Deteri- oration (PSD) Basic Information. Available at https://www.epa.gov/nsr/ prevention-significant-deterioration-basic-information (2025) 22

2025