pith. sign in

arxiv: 2602.16821 · v2 · submitted 2026-02-18 · 💻 cs.LG

TopoFlow: Topography-aware Pollutant Flow Learning for High-Resolution Air Quality Prediction

Ammar Kheder , Helmi Toropainen , Wenqing Peng , Samuel Ant\~ao , Jia Chen , Michael Boy , Zhi-Song Liu This is my paper

Pith reviewed 2026-05-15 21:10 UTC · model grok-4.3

classification 💻 cs.LG
keywords air quality predictionpollutant dynamicstopographywind directionvision transformerphysics-guided learningPM2.5high-resolution forecasting
0
0 comments X

The pith

A vision transformer that accounts for terrain shape and wind direction predicts air quality with substantially lower error than prior methods.

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

The paper introduces TopoFlow, a neural network designed to predict air pollutant levels at high resolution by explicitly incorporating how topography affects pollutant movement and how wind transports them. It modifies a standard vision transformer with topography-aware attention to capture terrain-induced patterns and wind-guided patch reordering to match spatial data to wind flows. This physics-guided approach is trained on extensive reanalysis data from China and delivers a PM2.5 root mean square error of 9.71 micrograms per cubic meter. That level represents major gains over both traditional operational systems and current AI models, while keeping errors below regulatory thresholds for reliable clean-versus-polluted distinctions.

Core claim

TopoFlow embeds physical processes into a neural network by using topography-aware attention to model terrain effects on flow and wind-guided patch reordering to align representations with wind directions. When trained on six years of high-resolution reanalysis data assimilating observations from over 1,400 stations across China, the model reaches a PM2.5 RMSE of 9.71 ug/m3. This is a 71-80 percent improvement over operational forecasting systems and a 13 percent improvement over state-of-the-art AI baselines. The advantage persists across all four major pollutants and forecast lead times from 12 to 96 hours, with errors staying below China's 75 ug/m3 air quality threshold.

What carries the argument

topography-aware attention and wind-guided patch reordering within a vision transformer that embed terrain channeling and wind transport into the model's spatial processing

Load-bearing premise

The two new mechanisms drive the gains more than any differences in data preparation or model tuning.

What would settle it

An ablation study that removes the topography-aware attention and wind-guided reordering from the model while keeping all other elements identical would show whether the RMSE increases significantly.

Figures

Figures reproduced from arXiv: 2602.16821 by Ammar Kheder, Helmi Toropainen, Jia Chen, Michael Boy, Samuel Ant\~ao, Wenqing Peng, Zhi-Song Liu.

Figure 1
Figure 1. Figure 1: TopoFlow architecture for physics-guided air quality prediction. The model takes as input concentrations of six air pollutants, major meteorological data, population density, spatial coordi￾nates, time stamps, and a topographic map, and outputs pollutant concentrations at lead times from 12 to 96 hours. All input data are stacked into a multi-layer 2D map, then cropped into non-overlapping patches. TopoFlo… view at source ↗
Figure 2
Figure 2. Figure 2: Overall performance of air pollution prediction. (a), Ground truth (CAQRA reanalysis) PM2.5 observations. (b), TopoFlow PM2.5 predictions. (c), Prediction error (|yˆ−y|, where yˆ is the model prediction and y the CAQRA reanalysis) distribution across lead times comparing TopoFlow, ClimaX, and AirCast. Box plots indicate median (middle line), 25th and 75th percentile (box), and 5th and 95th percentile (whis… view at source ↗
Figure 3
Figure 3. Figure 3: Forecast skill as a function of lead time for six air pollutants. RMSE validated against OpenAQ stations across China for 2019. a, PM2.5. b, PM10. c, NO2. d, SO2. e, CO. f, O3. TopoFlow (green) achieves the lowest errors for particulate matter and NO2. Aurora (purple) shows superior perfor￾mance for O3 and CO, which require three-dimensional atmospheric representation to capture stratospheric intrusions an… view at source ↗
Figure 4
Figure 4. Figure 4: Seasonal PM2.5 distribution from forecasts, reanalysis, and observations. (a–d), CAMS forecasts. (e–h), Aurora predictions. (i–l), CAQRA reanalysis. (m–p), TopoFlow predictions. (q–t), OpenAQ measurements. Columns: Winter (15 January 2019), Spring (1 March 2019), Summer (12 July 2019), Autumn (19 October 2019). TopoFlow achieves lowest RMSE (25.9 µg/m3 ) against inde￾pendent stations, outperforming CAQRA (… view at source ↗
Figure 5
Figure 5. Figure 5: Topographic blocking in Sichuan Basin. (a), CAQRA (ground truth) PM2.5 distribution across China at 7 July 2018, 20:00 UTC, with the study region marked by the green box. (b), CAQRA (ground truth) PM2.5 concentrations and wind vectors at forecast time (8 July 2018, 08:00 UTC) within the Sichuan Basin. The green line indicates 30.0°N transect. (c), TopoFlow 12-hour prediction achieving spatial correlation r… view at source ↗
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Physics-guided attention learns terrain-aware transport patterns. (a–c), Attention weight matrices over the Tianshan mountain region (elevation 26–5698 m): Baseline (a), Baseline + wind￾guided patch reordering (b), and Baseline + wind-guided patch reordering + elevation bias (TopoFlow) (c). Vertical stripes in (a) indicate shortcut-like routing through globally informative patches; wind￾guided reordering i… view at source ↗
Figure 8
Figure 8. Figure 8: Multi-pollutant model comparison over the Sichuan Basin (8 July 2018, 08:00 UTC, 12h forecast). Rows: PM2.5, PM10, SO2, NO2, CO, O3. Columns: CAQRA observations, TopoFlow, CAMS reanalysis, Aurora. All fields at 0.4◦ . R and RMSE computed against CAQRA. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
read the original abstract

We propose TopoFlow (Topography-aware pollutant Flow learning), a physics-guided neural network for efficient, high-resolution air quality prediction. To explicitly embed physical processes into the learning framework, we identify two critical factors governing pollutant dynamics: topography and wind direction. Complex terrain can channel, block, and trap pollutants, while wind acts as a primary driver of their transport and dispersion. Building on these insights, TopoFlow leverages a vision transformer architecture with two novel mechanisms: topography-aware attention, which explicitly models terrain-induced flow patterns, and wind-guided patch reordering, which aligns spatial representations with prevailing wind directions. Trained on six years of high-resolution reanalysis data assimilating observations from over 1,400 surface monitoring stations across China, TopoFlow achieves a PM2.5 RMSE of 9.71 ug/m3, representing a 71-80% improvement over operational forecasting systems and a 13% improvement over state-of-the-art AI baselines. Forecast errors remain well below China's 24-hour air quality threshold of 75 ug/m3 (GB 3095-2012), enabling reliable discrimination between clean and polluted conditions. These performance gains are consistent across all four major pollutants and forecast lead times from 12 to 96 hours, demonstrating that principled integration of physical knowledge into neural networks can fundamentally advance air quality prediction.

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.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes TopoFlow, a vision-transformer architecture for high-resolution air-quality forecasting that augments standard self-attention with a topography-aware attention module and replaces raster patch ordering with wind-guided reordering. Trained on six years of high-resolution reanalysis data assimilating observations from >1,400 Chinese stations, the model reports a PM2.5 RMSE of 9.71 µg m⁻³, corresponding to 71–80 % improvement over operational systems and 13 % over prior AI baselines, with consistent gains across four pollutants and 12–96 h lead times.

Significance. If the reported gains survive controlled ablation of the two novel modules, the work would provide concrete evidence that explicit encoding of terrain-induced flow and wind-driven transport can measurably improve neural forecasts in complex topography, offering a template for physics-informed architectures in other environmental domains.

major comments (2)
  1. [Experiments] Experiments section (and associated tables/figures): the manuscript presents no ablation experiments that remove topography-aware attention (reverting to standard ViT attention) or wind-guided patch reordering (reverting to raster order) while holding architecture depth, data volume, preprocessing, and training protocol fixed. Without these controls the 13 % margin over AI baselines cannot be causally attributed to the claimed inductive biases rather than differences in training data scale or hyper-parameter search.
  2. [§3.2 and §4] §3.2 and §4: the description of baseline implementations is insufficient to determine whether the reported 13 % improvement reflects a fair comparison. No details are given on whether the cited state-of-the-art AI models were re-trained on the identical six-year reanalysis corpus, used the same station density, or received equivalent hyper-parameter tuning.
minor comments (2)
  1. [Figure 3] Figure 3 caption and §4.1: the wind-direction histogram is shown without error bars or sample-size annotation, making it difficult to assess whether the reported alignment statistics are robust across seasons.
  2. [§3.1] Notation: the symbol for the topography embedding is introduced without an explicit equation; readers must infer its dimensionality from the attention diagram.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We agree that the claims would be strengthened by explicit ablation studies isolating the two novel modules and by expanded details on baseline implementations. We will revise the manuscript to incorporate both.

read point-by-point responses

  1. Referee: [Experiments] Experiments section (and associated tables/figures): the manuscript presents no ablation experiments that remove topography-aware attention (reverting to standard ViT attention) or wind-guided patch reordering (reverting to raster order) while holding architecture depth, data volume, preprocessing, and training protocol fixed. Without these controls the 13 % margin over AI baselines cannot be causally attributed to the claimed inductive biases rather than differences in training data scale or hyper-parameter search.

    Authors: We agree that controlled ablations are required to causally link the reported gains to the topography-aware attention and wind-guided reordering. In the revised manuscript we will add these experiments: variants using standard ViT self-attention (instead of topography-aware attention) and raster patch ordering (instead of wind-guided reordering), with architecture depth, data, preprocessing, and training protocol held fixed. Updated tables and figures will be included in the Experiments section. revision: yes

  2. Referee: [§3.2 and §4] §3.2 and §4: the description of baseline implementations is insufficient to determine whether the reported 13 % improvement reflects a fair comparison. No details are given on whether the cited state-of-the-art AI models were re-trained on the identical six-year reanalysis corpus, used the same station density, or received equivalent hyper-parameter tuning.

    Authors: We acknowledge that the current text lacks sufficient implementation details for the baselines. In the revision we will expand §§3.2 and 4 to state explicitly that all cited AI baselines were re-implemented and trained on the identical six-year reanalysis corpus with the same station density and preprocessing. We will also document the hyper-parameter search procedure used for each baseline to confirm the comparison is fair. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The manuscript presents an empirical neural network (vision transformer with two added mechanisms) trained on reanalysis data and evaluated on held-out forecasts. No equations, self-citations, or uniqueness theorems are invoked to derive the reported RMSE by construction from the inputs. The performance numbers (9.71 µg/m³, 13 % gain) are presented as measured outcomes of training and testing rather than tautological re-statements of fitted parameters or prior self-citations. The architecture choices are described as inductive biases motivated by domain knowledge, not as definitions that force the target metric. This is a standard empirical ML paper whose central claim remains falsifiable by independent replication or ablation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that topography and wind direction dominate pollutant transport at the scales considered, plus standard neural-network training assumptions. No free parameters or invented entities are declared in the abstract.

axioms (1)
  • domain assumption Topography and wind direction are the two critical factors governing pollutant dynamics
    Explicitly stated in the abstract as the physical basis for the two novel mechanisms.

pith-pipeline@v0.9.0 · 5565 in / 1190 out tokens · 16922 ms · 2026-05-15T21:10:40.599894+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages

  1. [1]

    State of Global Air 2024.Health Effects Institute Report, Boston, MA, 2024

    Health Effects Institute. State of Global Air 2024.Health Effects Institute Report, Boston, MA, 2024

    work page 2024

  2. [2]

    Nemery, B., Hoet, P. H. M., and Nemmar, A. The Meuse Valley fog of 1930: an air pollution disaster. The Lancet, 357(9257):704–708, 2001. 27 TopoFlow: Topography-aware Pollutant Flow Learning for High-Resolution Air Quality Prediction

    work page 1930

  3. [3]

    Bei, N. et al. Typical synoptic situations and their impacts on the wintertime air pollution in the Guanzhong basin, China.Atmospheric Chemistry and Physics, 16:7373–7387, 2016

    work page 2016

  4. [4]

    Wang, Y. et al. Aggravated air pollution events and high secondary aerosol formation potential in the Sichuan Basin, China.Journal of Geophysical Research: Atmospheres, 123:4292–4308, 2018

    work page 2018

  5. [5]

    D.Mountain Meteorology: Fundamentals and Applications

    Whiteman, C. D.Mountain Meteorology: Fundamentals and Applications. Oxford University Press, 2000

    work page 2000

  6. [6]

    Ge, C. et al. Validation and calibration of CAMS PM2.5 forecasts using in situ PM2.5 measurements in China and United States.Remote Sensing, 12(22):3813, 2020

    work page 2020

  7. [7]

    Ambient Air Quality Stan- dards

    Ministry of Environmental Protection of the People’s Republic of China. Ambient Air Quality Stan- dards. GB 3095-2012, Beijing, China, 2012

    work page 2012

  8. [8]

    A Correction Method of Environmental Meteorological Model Based on Long-Short-Term Memory Neural Network.Earth and Space Science, 6(12):2423– 2434, 2019

    Dai, H., Huang, G., Zeng, H., and Yang, F. A Correction Method of Environmental Meteorological Model Based on Long-Short-Term Memory Neural Network.Earth and Space Science, 6(12):2423– 2434, 2019

    work page 2019

  9. [9]

    Qi, H. et al. Multi-model Evaluation and Bayesian Model Averaging in Quantitative Air Quality Forecasting in Central China.Aerosol and Air Quality Research, 22(3):210247, 2022

    work page 2022

  10. [10]

    Ali, M. A. et al. Accuracy assessment of CAMS and MERRA-2 reanalysis PM2.5 and PM10 concen- trations over China.Atmospheric Environment, 288:119297, 2022

    work page 2022

  11. [11]

    Jiang, Y. et al. Contribution of Regional PM2.5 Transport to Air Pollution Enhanced by Sub-Basin Topography: A Modeling Case over Central China.Atmosphere, 11(11):1258, 2020

    work page 2020

  12. [12]

    Improving WRF-Chem PM2.5 predictions by combining data assimi- lation and deep-learning-based bias correction.Environment International, 195:109186, 2024

    Peng, Z., Liu, H., and Ma, X. Improving WRF-Chem PM2.5 predictions by combining data assimi- lation and deep-learning-based bias correction.Environment International, 195:109186, 2024

    work page 2024

  13. [13]

    K., and Grover, A

    Nguyen, T., Brandstetter, J., Kapoor, A., Gupta, J. K., and Grover, A. ClimaX: a foundation model for weather and climate. InInternational Conference on Machine Learning, 2023

    work page 2023

  14. [14]

    Lam, R. et al. Learning skillful medium-range global weather forecasting.Science, 382(6677):1416– 1421, 2023

    work page 2023

  15. [15]

    Bi, K. et al. Accurate medium-range global weather forecasting with 3D neural networks.Nature, 619(7970):533–538, 2023

    work page 2023

  16. [16]

    Bodnar, C. et al. A foundation model for the Earth system.Nature, 641:1180–1181, 2025

    work page 2025

  17. [17]

    Dosovitskiy, A. et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. InInternational Conference on Learning Representations, 2021

    work page 2021

  18. [18]

    Liang, J. et al. SwinIR: Image Restoration Using Swin Transformer. InIEEE/CVF ICCV Workshops, pages 1833–1844, 2021

    work page 2021

  19. [19]

    Vaswani, A. et al. Attention is all you need. InNeurIPS, pages 6000–6010, 2017

    work page 2017

  20. [20]

    Hettige, K. H. et al. AirPhyNet: Harnessing physics-guided neural networks for air quality prediction. arXiv preprint arXiv:2402.03784, 2024

    work page arXiv 2024

  21. [21]

    Xu, Y. et al. Improving air quality assessment using physics-inspired deep graph learning.npj Climate and Atmospheric Science, 6:152, 2023. 28 TopoFlow: Topography-aware Pollutant Flow Learning for High-Resolution Air Quality Prediction

    work page 2023

  22. [22]

    Deep Spatio-Temporal Neural Network for Air Quality Reanalysis

    Kheder, A., Foreback, B., Wang, L., Liu, Z.-S., and Boy, M. Deep Spatio-Temporal Neural Network for Air Quality Reanalysis. In Petersen, J. and Dahl, V. A., editors,Image Analysis (SCIA 2025), pages 74–87. Springer, 2025

    work page 2025

  23. [23]

    Nedungadi, V. et al. AirCast: Improving Air Pollution Forecasting Through Multi-Variable Data Alignment. InICML 2025 Workshop, 2025

    work page 2025

  24. [24]

    Kong, L. et al. A 6-year-long (2013–2018) high-resolution air quality reanalysis dataset in China based on the assimilation of surface observations from CNEMC.Earth System Science Data, 13(2):529–570, 2021

    work page 2013

  25. [25]

    NOAA Technical Memorandum NESDIS NGDC-24, 2009

    Amante, C.andEakins, B.W.ETOPO11Arc-MinuteGlobalReliefModel: Procedures, DataSources and Analysis. NOAA Technical Memorandum NESDIS NGDC-24, 2009

    work page 2009

  26. [26]

    Gridded Population of the World, Version 4 (GPWv4): Population Density, Revision 11

    Center for International Earth Science Information Network (CIESIN). Gridded Population of the World, Version 4 (GPWv4): Population Density, Revision 11. NASA SEDAC, 2018

    work page 2018

  27. [27]

    OpenAQ – Open Air Quality Data.https://openaq.org, 2024

    OpenAQ. OpenAQ – Open Air Quality Data.https://openaq.org, 2024

    work page 2024

  28. [28]

    M.Pattern Recognition and Machine Learning

    Bishop, C. M.Pattern Recognition and Machine Learning. Springer, New York, 2006

    work page 2006

  29. [29]

    Inness, A. et al. The CAMS reanalysis of atmospheric composition.Atmospheric Chemistry and Physics, 19(6):3515–3556, 2019

    work page 2019

  30. [30]

    Validation and Calibration of CAMS PM2.5 Forecasts Using In Situ PM2.5 Measurements in China and United States.Remote Sensing, 12(22):3813, 2020

    Wu, C., Li, K., and Bai, K. Validation and Calibration of CAMS PM2.5 Forecasts Using In Situ PM2.5 Measurements in China and United States.Remote Sensing, 12(22):3813, 2020

    work page 2020

  31. [31]

    Lin, M. et al. Climate variability modulates western US ozone air quality in spring via deep strato- spheric intrusions.Nature Communications, 6:7105, 2015

    work page 2015

  32. [32]

    Knowland, K. E. et al. Stratospheric intrusion-influenced ozone air quality exceedances investigated in the NASA MERRA-2 reanalysis.Geophysical Research Letters, 44(20):10691–10701, 2017

    work page 2017

  33. [33]

    Zhao, D. et al. Impact of stratospheric intrusions on ozone enhancement in the lower troposphere and implication to air quality in Hong Kong and other South China regions.Journal of Geophysical Research: Atmospheres, 126(3):e2020JD033955, 2021

    work page 2021

  34. [34]

    Lefohn, A. S. et al. Quantifying the importance of stratospheric-tropospheric transport on surface ozone concentrations at high- and low-elevation monitoring sites in the United States.Atmospheric Environment, 62:646–656, 2012

    work page 2012

  35. [35]

    Kar, J. et al. Evidence of vertical transport of carbon monoxide from MOPITT.Geophysical Research Letters, 31(23):L23105, 2004

    work page 2004

  36. [36]

    Park, M. et al. Transport pathways of carbon monoxide in the Asian summer monsoon diagnosed from MOZART.Journal of Geophysical Research: Atmospheres, 114(D8):D08303, 2009

    work page 2009

  37. [37]

    Foreback, B. et al. A Lagrangian view on severe haze in Beijing: local and long-range sources of trace gases and primary and secondary aerosols.Atmospheric Environment, 363:121602, 2025

    work page 2025

  38. [38]

    Xu, H. et al. Permutation Equivariance of Transformers and its Applications. InIEEE/CVF CVPR, pages 5987–5996, 2024

    work page 2024

  39. [39]

    Air Pollution, the Automobile, and Public Health.Environment: Science and Policy for Sustainable Development, 31(4), 1989

    Wilson, R. Air Pollution, the Automobile, and Public Health.Environment: Science and Policy for Sustainable Development, 31(4), 1989. 29

    work page 1989