General Comments

This paper has the primary goal of presenting a dataset combining several long-term metrics related to surface ozone concentration together with explanatory parameters (metadata) related to the ozone measurement stations, which can be used as inputs to a machine learning framework to predict the metrics. It also presents a benchmark application of several simple machine learning algorithms to this problem as an illustrative example. Overall the dataset and its goal are very useful, but I believe several improvements should be made prior to its publication.

The goal (as summarized best, I think, at the start of Section 4) is to produce estimates of various long-term ozone metrics based on local site conditions, rather than (for example) short-term estimates based on atmospheric characteristics (temperature, wind, etc.). I think this should be more clearly and explicitly stated up-front, ideally in the abstract, so the specific goals for which this dataset has been created are more immediately clear to a reader.

Although it is perhaps beyond the scope of this work, I believe that the introduction should also make mention of the use of machine learning in other areas related to atmospheric chemistry and air quality. Machine learning techniques are often used as emulators or surrogate models for more computationally complex components of atmospheric chemistry models, i.e., to replace complicated and costly atmospheric chemistry calculations with simpler machine learning surrogates to improve computational performance of the models. Second, machine learning is being extensively used in the calibration of low-cost sensors for air quality, in order to account for the many sources of interference with the measurements of these sensors and allow them to more effectively supplement the existing monitoring networks for pollutants such as ozone. Although such use of machine learning is beyond the scope of the present work, I believe it is still worthwhile to draw attention to these other areas of ongoing work combining machine learning and atmospheric science.

In Table 1, sources should be provided for these metadata, especially data on NOx emissions, NOx column, and night light intensity. It should also be made more clear what time period these represent, i.e., are they the average over the entire time interval, or more representative of the situation in one particular year. I suspect that the answer will vary by dataset; for example, population density might be based on census data from a particular year, while NOx column density might be derived from a satellite, and therefore represent an average (although only for clear-sky periods during which the satellite is overhead) over many years. Details do not necessarily need to be provided for each underlying dataset, but the initial source for each data should be made clear so the reader can look into the details.

In Table 2, there seems to be an inconsistent definition for data completeness applied across these metrics (e.g., the overall average requires only one valid datapoint, percentiles require 10, and daytime and nighttime averages require 75% completeness). I would suggest, at a minimum, that a “completeness” metric also be included in the data set, indicating what fraction of the expected total number of hourly measurements for each site were actually present in the dataset. This would allow users to potentially make different decisions about what values they consider to be “valid” to their applications. Furthermore, if the datasets are more complete towards the beginning or end of the time period in question, this could combine with the fact that the underlying metadata (e.g., population density) might be changing over time, and be a confounding factor in the performance of the machine learning approaches. While this is a complex issue to address, some mention should be made of this limitation on the data which a single “completeness” metric will not capture.

In section 5.1, a particular machine learning task and evaluation metric is introduced. While this is useful for benchmarking purposes, it may not be appropriate in all cases, depending on the application. For example, determination of whether metrics fall into certain discrete classification regimes, e.g., “healthy” or “unhealthy”, may be a desired goal in certain applications. While it is impossible to account for all potential uses of these metrics, some mention should be made that this is only one of many possible machine learning goals and evaluation metrics.

Specific Comments

Line 36: O3 and Ozone are both used; this may be redundant.

Line 79: Starting the sentence with “I.e.” seems awkward to me; I would suggest rephrasing this.

Lines 105-106: suggest revising to “Ozone irreversibly damages plant tissue when the plant leaves take it up (Schraudner et al., 1997) leading to reduced crop yields (Mills et al., 2011).”

Line 110: suggest removing “exemplary”.

Lines 127-128: This sentence is awkward; I am not sure what the correct phrasing should be, but the authors should consider revising this sentence.

Line 132: suggest revising “ground-level ozone levels” to “ground-level ozone concentrations” to avoid the repeated word.

Line 169: It is unclear whether “mean ozone metrics” refers to the average values of different metrics across the given time period, or simply the average ozone concentrations in different locations. This may need to be clarified.

Lines 233-236: It is unclear how these clusters are used. Are entire clusters assigned to one of the three data subsets, or are individual members of the clusters divided between subsets, such that each subset will get at least one sample from each cluster? Also, how is the issue of sparse measurements in South America or Africa, for example, addressed using this spatial clustering approach?