This work developed a new approach for hourly CO2 emission mapping at high resolution from on-road traffics for 20 cities in France, Germany, and the Netherlands in 2023, with the FCD data created from GPS information, the traffic volume data based on machine leaning models, and speed- and vehicle type-specific emission factors. The new CO2 emissions from on-road transportation were validated and the spatial and temporal variation characteristics were presented and discussed. The manuscript is generally well written. There are some comments which required to be addressed before it can be accepted.

1. Pay attention to the blank before the bracket particularly in the Introduction Section.
2. Line 13, point out the CO2 emission from on-road traffic or transportation.
3. Line 44-55, the details of this paragraph are not so necessary. Simplify the sentences and link them more to the major contents of this study.
4. Line 121, The title of Figure 1 is not correct.
5. Line 165, in Table 2, for the road-specific traffic count data, are they daily or hourly? Only hourly traffic volume can be used to produce hourly emissions.
6. Line 195-196, why use monthly average instead of hourly or daily average meteorological data to calculate the emission factors?
7. Line 206-207, clarify the potential uncertainty caused by using a standard EFCO2, instead of a measured EFCO2 from literature.
8. Line 345-346, this sentence is repeated.
9. Line 389-390, Is there any difference in the emission factors used for calculations which could cause the discrepancies?

This paper is interesting and necessary for quantifying urban carbon emissions. I'm impressed by the authors' effort to present a high-resolution approach for road traffic CO2 emissions. This kind of information is very important  for urban planning, both now and in the future. In fact, this work tackles a necessary issue in urban emission monitoring: representing the emissions of individual roads across an entire city scale. The study has many strengths: the claim of providing the first hourly, street-level emissions for 20 European cities is a major achievement, even if the "first time" assertion might be too strong (other studies have done similar work, though not with this many cities). 

Furthermore, the methodology is well-structured and transferable, with solid documentation of the data processing steps. The applied use of ML and the choice of the algorithm are, in my view, well-justified. Finally, it covers a considerable spatial extent, including small roads ( e.g., residential classe) that are often omitted from many inventories and even high-definition urban traffic emission studies. 

Although I acknowledge the study's importance, several methodological concerns and uncertainties must be addressed by the authors before publication. My review is divided into the following points:

1 _____

The most critical weakness of this study based on the extrapolation/generalization of the GPS-to-volume conversion and the resulting CO2 emissions, since this decision uses a ML model trained exclusively on data from Paris and Berlin and applies it to the remaining 18 cities. While the authors justify this by stating that "high-quality in-situ traffic observations are either unavailable or not publicly accessible for other cities," this extrapolation represents a major source of uncertainty that needs better justification and thorough discussion.

The main issue is the assumption that factors, such as GPS penetration rates, fleet compositions, and traffic behavior, in cities like Munich, Amsterdam, and Lyon will match those in Paris or Berlin. This assumption may hold true for some roads but is unlikely to be valid across the board. This simplified idea has strong impacts on the results and their applicability in urban contexts.

Have look at this: the ML model validation showed poor performance on middle and small roads in Paris (car R2 0.33, truck 0.23) and major in Berlin (car 0.66, no truck). I suppose, applying this model to other cities with potentially different GPS penetration rates and urban structures introduces unquantified errors, particularly on those less-traveled routes. Importantly, no sensitivity analysis is provided by authors to quantify how variations in GPS penetration (and, consequently, traffic volume estimates) affect the final emission estimates. To do so, I recommend to perform a sensitivity analysis demonstrating how hypothetical variations in GPS penetration (e.g., ±10%, ±20%, ±30% ) affect the total estimated emissions for the extrapolated cities; provide explicit uncertainty bounds for each extrapolated city, reflecting the potential error introduced by model transferability. This also leads me to second point.

2___

I am not convinced that validation strategy, limited to 2 cities-  Paris and Berlin, is sufficient for main claim made about all 20 cities. Note that no independent validation for ~90% of the cities analysed. 

I am not sure it if is possible, but I would suggest an external validation to compare the model’s volume estimates with any available traffic statistics (even if only annual or from limited sites) from the 18 extrapolated cities. 

Please, clarify the state of calibration in the abstract and conclusions, explicating the emission models for 18 of the 20 cities utilize uncalibrated, extrapolated models. 

There are also other related-concerns:

Comparison with Carbon Monitor (Figure 6) shows moderate correlations (R = 0.58-0.84) but systematic differences are not adequately explained.  It is important to note that CM-city estimates are also based, in part, on consumer-driven mobility data like TomTom GPS. While Floating Car Data (FCD), such as that from TomTom, is valuable, it introduces significant discrepancies when compared against local traffic flow, as noted in previous literature (e.g.,doi:10.1002/essoar.10504783.1, doi:10.5194/egusphere-egu21-5419). The large discrepancies observed here warrant a much deeper investigation than a simple attribution to general "methodological differences.

The large discrepancies when comparing annual estimates with other high-resolution studies and city-specific inventories (ranging from −94% to −8%, and showing a ∼80% difference in Table S7) are concerning and must be better explained. The authors need to explain the major differences compared to the following studies: Ulrich et al., 2023 (-8.1% - low difference), Anjos et al., 2025 (-66 %) and Kühbacher et al.,2023 (-74.2%). It is important  to note that the studies by Kühbacher et al. (which uses a bottom-up traffic model like VISUM) and Anjos et al. (which uses an ML-based bottom-up approach) both rely on local traffic counts from monitoring stations for their volume inputs.

My question is: What factor (s) is (are) limiting the CO2 emission estimates derived from the FCD-based ML model? Given the inherent discrepancies in FCD when estimating actual traffic volume, is there a systematic bias in the GPS to volume conversion that consistently leads to the underestimation of emissions compared to inventories and studies that are anchored to local traffic counts?

3____

The choice to simply adopt an 80% training and 20% testing split can be quite simple in the context of ML. This "naïve" splitting method can not be fully minimize overfitting or ensure the model is robust and generalizable to new, unseen data (data outside this study's scope). Why didn't you consider the validation techniques such as k-fold cross-validation, chronological splitting for time series data, or bootstrapping? These techniques are widely used to evaluate both the gap-filling model and the GPS-to-volume conversion ML model.

Furthermore, while R2 is a good metric for assessing fit, it doesn't measure the error magnitude itself. To provide a complete picture of model performance, you should include RMSE and MAE from Table S2 directly in the main text.

Table 2 lists eight features, but the justification for choosing these specific features is limited. Please provide a clearer explanation, including a literature basis, for why these predictors were selected over others. Since more potential predictors can often improve ML model performance, testing different feature engineering approaches for different road classes could be addressed for future research.

4___

It's great that the data is available on Zenodo. But, the code is not mentioned. Since the Python code is integral to your entire methodology—covering all steps, from pre-processing and spatial operations to training, prediction, and CO2 emission calculation—I strongly recommend depositing it on GitHub and Zenodo. Providing well-documented code and samples is necessary for transparency and reproducibility.

5____

The paper states "we used EFCO₂ of the EU6 standard" because "CO₂ emission factors are only marginally influenced by emission standards" (p.10). I think is imprecise due to Euro standards primarily target air pollutants, but CO₂ varies also significantly by vehicle age/technology, for instance. 

The final urban CO2 emissions are currently based on 100x100 meter grid cells. For better comparability and utility in urban modeling, the emissions data, in general, are calculated and expressed as a density per unit area within those grid cells. Please, check if it is more suitable reporting the CO2 emissions in units of mass per area per time (e.g., CO2 m2 ) rather than in units derived from line-segment meters.

6__

While uncertainties are discussed qualitatively, no quantitative uncertainty estimates are provided for the emission maps. Note that: ML models have associated R² values, but these are not propagated to final emission uncertainties; GPS data coverage varies dramatically depending on road class, but impact on final emissions is not quantified. All of these limitations lead to a cumulative effect of multiple uncertainty sources that is unknown and unreported in the text.  

It will be useful, at least, provide uncertainty ranges for annual city emissions (e.g., Berlin: 7.9 ± X Mt CO₂,  Paris: 1,94 ± X Mt CO₂). 

Please, be carful to clearly distinguish between precision (model R²) and accuracy (comparison with true emissions)

7__

The limitations are well-discussed in Section 4 but should be elevated, in my opinion. Maybe adding brief limitations statement to abstract, creating a "Limitations" subsection in discussion, and quantifying limitations wherever possible (don't just say "may lead to bias”).