the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Estimating Hourly Surface NO₂ Across the Contiguous United States Using TEMPO Observations and Machine Learning
Abstract. Ground-level nitrogen dioxide (NO2) poses significant health risks, yet hourly concentrations have not been estimated due to the limited temporal sampling of polar-orbiting satellites. NASA’s Tropospheric Emissions: Monitoring of Pollution (TEMPO) geostationary instrument provides the first hourly tropospheric NO2 observations over North America, creating new opportunities for estimating surface NO2 throughout the daytime. We develop XGBoost models using TEMPO tropospheric NO2, ERA5 meteorology, road density, impervious surface percentage (ISP), and temporal features to predict hourly surface NO2 at 474 EPA stations across the contiguous United States from August 2023 through June 2025, and map concentrations at TEMPO’s L3 resolution. Separate TEMPO (daylight) and no-TEMPO (nighttime/imputation) models achieve test R2 = 0.78 (RMSE = 3.33 ppb) and R2 = 0.76 (RMSE = 4.36 ppb), respectively. TEMPO column NO2 is the strongest predictor by gain, while ISP dominates spatial prediction and boundary layer height is the most influential meteorological predictor. Importantly, TEMPO provides complementary spatial information beyond meteorological and temporal predictors, increasing R2 across all spatial cross-validation folds by an average of 17% and improving performance during data-scarce periods. Spatial cross-validation (R2 = 0.48) further highlights opportunities to improve predictions at unseen locations through higher-resolution spatial inputs, while extending satellite coverage into the evening could improve representation of the evening NO2 peak. To our knowledge, this is the first study to predict hourly surface NO2 across CONUS using geostationary satellite data, demonstrating the potential of TEMPO to support acute exposure assessments at previously infeasible temporal resolutions.
- Preprint
(3780 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 04 Dec 2026)
- RC1: 'Comment on egusphere-2026-6352', Anonymous Referee #1, 17 Nov 2026 reply
-
RC2: 'Comment on egusphere-2026-6352', Anonymous Referee #2, 19 Nov 2026
reply
This manuscript describes a methodology to estimate hourly surface NO2 concentrations from TEMPO observations. While a product like this is needed in the community, the authors have several key methodological shortcomings. For that reason, I am recommending substantial revisions, with final publication in doubt.
Major comment #1:
It appears the authors are retaining all cloudy retrievals (in addition to the clear-sky retrievals). Cloudy data are typically excluded for two primary reasons. First, because the satellite cannot “see” through optically thick clouds. So a measurement on a cloudy day is the amount of NO2 between the sensor and top of the optically thick cloud. The assumptions in averaging kernel can sometimes make these cloudy measurements look reasonable, but these hardly constitute true measurements. Please re-read literature on this topic. Second, the uncertainty is much higher when there is partial cloudiness, especially around cloud fractions of 0.2 – 0.5 is very high, and thus the TEMPO Users’ Guide strongly recommends not using this data for quantitative purposes.
Instead the best and most common practice is to impute the missing data instead of using the cloudy data. Most often this information comes from a chemical transport model. Or it could be imputed from the surface observations themselves.
Nonetheless, I am intrigued by the approach because although the cloudy data has substantially more errors, there is some usable information if you use it properly. The authors need to better substantiate why they are using cloudy data as this is not in alignment with the current scientific literature. I think the best way forward would be to train a model using the appropriate cloud filter recommendations in the TEMPO Users’ Guide and compare with their method of retaining the cloudy data. I could imagine both would have their advantages and disadvantages, but again, solely including a model with cloudy is inappropriate because this is not in alignment with previous literature.
Major comment #2:
I am concerned that your “no TEMPO” model has such good performance. To me, this suggests that the model is substantially overfit, even if your statistics say otherwise. I am particularly concerned about predictions away from the monitors. The Permian Basin in Texas is an example. To my knowledge, there are no NO2 monitors in this region, yet it is a NO2 hotspot picked up only by TEMPO and none of your other datasets. The same can be said for rural power plants, wildfires, agricultural/soil NOx, etc. Not including TEMPO will get NO2 predictions away from monitors for the wrong reasons. If the authors had included NO2 from a CTM as a predictor that would be a different story.
I suppose an argument could be made that including a “no TEMPO” model as an interesting case study. But all it shows is a model overfit for the wrong reasons, and therefore scientifically has little value for NO2 applications away from a monitoring location.
A better way would have been to include TEMPO NO2 measurements from the previous daytime to infer nighttime estimates. This would have been OK, but excluding TEMPO altogether at night, is not reasonable.
Minor comments:
Lines 29 – 35. I can tell these sentences have been written by AI, partially because AI uses slightly different subscript formation (which the authors did not catch). This is, in theory, OK (to me as a reviewer, but I cannot speak on behalf of the journal) as nothing is factually wrong, but it is dense and nuance is missing how it is relevant to this particular project. It reads like a general summary of NO2 vertical gradients as if someone not involved in the project would write it. I am sure I am only catching a fraction of the text partially written by AI. It comes off as dense and disjointed, and is not a good look. I would recommend going through and personally re-writing this and other sections (Lines 106 -112 is another example) that were written, in part, by AI.
Lines 82 – 87. Please mention the general downsides of ML. In particular, most ML models do not capture the extreme events well, because their training dataset may not have these extreme events adequately captured or captured at all. For example, a ML model might be trained by a monitor in a polluted city, but not at the precise address of the worst pollution. Therefore, when trying to predict at the worst address it will underestimate. Whereas CTMs do not suffer from this. If you provide a CTM with initial conditions that produce an extreme event, they will forecast an extreme event. Additionally, you allude in the Abstract about spatial overfitting, and that would be helpful to mention here too.
Line 93. Remove this sentence. Although technically true upon submission, there are other manuscripts under review and this might not actually be the first one published.
Lines 132-146. The authors have missed the fact that some NO2 monitoring sites are denoted as near-road sites. These sites are located in urban, suburban, and rural areas, and have unique characteristics. These monitors are located usually 10 m from major 8-lane highways and are not representative of 1 km grid boxes. Therefore, the authors will need to re-think this classification as they are missing near-road sites as a classification. I would recommend flagging these sites and potentially not using them, if you would like accurate 1 km predictions. The ISP variable that the authors use (Line 218) is insufficient. These near-road monitors are located near 8-lane major highways, not any “impervious surface”.
Line 138. It drops on that date because it was not released yet in the AQS database, not necessarily that the data don’t exist or that something nefarious was going on. Please clarify.
Line 146. I think you mean retain “only” the CAPS. Right?
Line 153. Local standard time? Or local time accounting for DST and standard time?
Line 166. That’s not the correct doi for the TEMPO data (just Earthdata in general, which is not fully relevant here). Please use the correct doi.
Lines 272 – 277. The discussion here makes me uncomfortable. What are the geophysical reasons behind a variable being correlated or not. For example, stratospheric NO2 is known to have no correlation with the surface NO2 (so why was this even included as an option in the first place?) For other variables, such as solar zenith angle, from a geophysical perspective, I would have thought it would improve correlation. It insufficient to simply say inclusion of a variable decreases correlation and move on, especially when it contrasts geophysically with what one might expect.
Figure 8 is very helpful. Although it does lead to some unanswered questions. First, why is the TEMPO 17:00 local time panel missing? The bottom right panel is not helpful. Now focusing on the top row, it is hard to tell which locations have the largest NO2 values, but there seem to be random single grid cells with red shading next to grid cells that are green or yellow. This would make some sense if it were a power plant or large source, but not too many of these exist in this region. Instead, I would have expected NO2 to show more dramatic enhancements near major highways and the downtown city center, which also aligns with the second row. Or in other words, some of the spatial gradients seem off. My guess is that there is some aspect of overfitting at play here, since all monitors in this image are heavily skewed to being in high population density zones. This once again points to some methodological shortcomings (already discussed in prior comments). A methodological work-around could have been to more heavily weight rural monitors when training the model.
Sections 3 and 4 are well-written, and I have few comments here. The analysis of the model is great. The problem with this manuscript lies in the methodology. My one suggested comment with Section 3 is to shorten the text at your own discretion. The figures sometimes speak for themselves.
Citation: https://doi.org/10.5194/egusphere-2026-6352-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 353 | 132 | 24 | 509 | 35 | 30 |
- HTML: 353
- PDF: 132
- XML: 24
- Total: 509
- BibTeX: 35
- EndNote: 30
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
In “Estimating Hourly Surface NO2 Across the Contiguous United States Using TEMPO Observations and Machine Learning” the authors train a gradient boosted regression tree model using observations of NO2 from EPA monitors as the training target and meteorological data as well as observations from TEMPO as the inputs. The authors evaluate model performance over different times of day, degrees of urbanization, and region. In general, the model captures observed variability in NO2, although it tends to underestimate NO2 at the highest range of the distribution and fails to capture the variability at sites (e.g., those in rural areas) with low concentrations. The authors thoroughly discuss limitations of the approach. While this is an interesting paper that highlights the value of TEMPO observations in constraining surface NO2, several issues need to be addressed before the paper is suitable for publication.
Major Concerns:
To increase data availability and coverage, the authors use observations from TEMPO that are highly uncertain, including pixels with cloud fraction above the recommended threshold and pixels at high solar zenith angle. While the authors do acknowledge the uncertainty associated with these data and include a metric for uncertainty in their ML model, there needs to be more quantitative discussion about the impact of the inclusion of these data on the product accuracy and in the uncertainty of the resultant NO2 product. For example, how does the performance of this model compare to one using only the highest quality TEMPO data (e.g., cloud fraction < 20%, SZA < 70 deg, quality flag = 0)? Is the accuracy of the model higher for grid boxes with the higher quality data? How does the model handle the highly uncertain data (e.g., does it ignore the TEMPO data if it is highly uncertain?)?
Related to this concern, it would be helpful to quantify the uncertainties in your final product, or at least to discuss the need to quantify these results. How do the uncertainties in the NO2 retrieval and the in situ NO2 observations translate to uncertainty in the final surface NO2 product? If the ultimate goal is helping to inform health studies, then at least showing that it is feasible to determine quantifiable uncertainties should be a priority.
Specific Concerns:
Line 45: Is the higher mortality rate associated with NO2 exposure or in general? Please clarify.
Line 145: While you use only CAPS instruments at sites that have both CAPS and chemiluminescence monitors, how does the inclusion of the mixture of CAPS and chemiluminescence from different sites affect your results? Does the difference in detection limits and uncertainties affect model accuracy at all, particular at rural sites with lower concentrations?
Line 166: Version 3 data are not the latest data. Version 4 data are available and are significantly different in some regions (e.g., those with large changes in elevation). While it’s not necessary to redo the entire analysis with the V4 data, acknowledging potential changes and effects on the product is warranted, as well as changing the phrasing in this line.
Line 171: Does this mean you are including quality flags with value 0 and 1, and just removing data with quality flag = 2? Data with QF = 1 are still highly uncertain, as they include observations with high SZA that have limited sensitivity near the surface (I believe you mention this later), so they are likely more representative of the a priori.
Line 185: Similar question as above for the inclusion of the data with high cloud fraction. How does the high uncertainty of these data affect the results? Do you see better accuracy for lower cloud fraction data?
Section 2.2.3: Why are you using such coarse resolution meteorological data if your goal is to resolve NO2 at the TEMPO resolution? Why not use something like HRRR that is available at approximately the spatial scale of a TEMPO pixel? I would imagine the reanalysis data cannot capture intra-urban variations in meteorology given its relatively large scale.
Line 321: Having a supplementary figure with the 10 regions would be helpful.
Line 386: When you say “strongest agreement” what are you referring to? Strongest compared to what? The stratospheric portion?
Line 394: It’s true that including the high SZA data increases the number of observations that you have, but doesn’t it potentially degrade your model performance by including “bad” data?
Line 396: I would recommend removing this paragraph. There is no expectation that stratospheric NO2 column should be correlated with surface NO2. The paper is long as it is, so removing unnecessary information, like this, would increase readability.
Line 431: There are techniques to try to increase performance at the high end of a distribution (e.g., resampling the dataset to create a different distribution, oversampling data). Did you try any of these? It seems like accuracy at these high values should be a priority, since your focus is on health applications.
Figure 4: While your figure is fine as is, the usual format is to have the “truth” on the x-axis and the modeled/predicted value on the y-axis. I would recommend presenting the data in this format.
Line 465: I’m not following the discussion here. Shouldn’t r2 values equal the square of the Pearson’s correlation coefficient (r)? What are you comparing here that results in this not being true?
Figure 5: How representative are the data in this figure? You’re just showing one week for each season. Is this particular week representative of all weeks? How does this vary within a season? Is the general relationship among sites always true? You don’t necessarily need to show data for all weeks, but some discussion of this variability is warranted.
Figure 8: On the middle row, it could be useful to show an indicator of urbanization on the map, so that people who are less familiar with Northeastern US geography can follow along with the discussion. Also, for the panel showing the time series, it would be better if you could format the lines so that the reader can match the actual and predicted values for a specific location.
Tables 2 & 3: I found these confusing. Formatting the tables so that it’s clear that you’re showing three different CV approaches would greatly improve readability. For example, including the End Date for the Temporal CV approach makes it look like those dates apply to the Spatial columns as well. For Table 3, labelling your columns as ΔRMSE and ΔR2 would also improve readability.
Discussion: I think mentioning the applicability of this methodology to other geostationary satellites, like Sentinel 4, is warranted given that this is a European journal.