The study is oriented towards identifying airborne pollen and the authors imply that flow cytometry is an efficient alternative for microscopical identification. It is particularly valuable to see that analysis of flow cytometry measurements seems to enable discrimination between different species of the same plant genus which is usually not possible in routine microscopical analysis. The manuscript is well written, and it clearly describes possibilities of standard flow cytometers for identification pollen.

However, in my opinion the manuscript lacks tests and discussion on the applicability of proposed method (flow cytometry measurement and the developed random forest classification model) for analysing aerosol samples. The authors emphasized importance of adapting the models to real environment samples (lines 259-264). But in my opinion for Atmospheric Measurement Techniques more than just theoretical discussion is needed when linking to atmospheric measurements. Without tests on atmospheric samples, it is just a speculation that proposed approach has a “potential for large-scale urban pollen monitoring”.

There are several aspects that should be addressed/discussed:

1. How the identification algorithm handles large diversity of aerosols (i.e. fungal spores, dust clusters, starch) which could be in the same size class as pollen. So far, the studies showed that very good classification algorithms in lab settings (as seen in confusion matrix) in real environment tend to fail.
2. The authors claim using flow cytometers could enable near-real-time pollen identification. How the aerosol sample is to be processed to be delivered to the flow cytometer (here bear in mind that sample could contain large particles that might require filtering out to avoid clogging).
3. Also related to previous item please discuss to what extent the flow cytometry approach could be challenging for quantification. For example, some devices can handle up to 200 microliters of sample for a single analysis which then requires either subsampling from larger volumes in which aerosols are collected or sample should be concentrated.

The approach to rely pollen identification exclusively on flow cytometry measurements that most cytometers routinely used in healthcare is very important. But the use of the same classification algorithm on different devices (even the same model) appeared to be challenging (as authors also clearly noted in lines 273-281). If not possible to test the model on different device measuring same parameters, the authors should at least discuss the measurement uncertainty for each parameter and refer to other studies that observed differences in flow cytometry parameters between different devices.

In line 133 authors indicated the training dataset for Thuja genus was impossible to clean from debris. Is presence of debris confirmed by microscope? If not, how can you be sure it is not a part of the normal pollen variability? The pollen from Thuja (and many other Cupressaceae) tends to break in wet environment resulting in separation of exine from the resto of pollen grain. Could it be that those separated exines are the “debris” you see in the data.

In Table A1, authors reference is missing for an accurate scientific name (e.g. Ambrosia artemisiifolia L.). Genus should be written in cursive font and also should include author references

Overview:

In this study, Tardiff et al. sought to distinguish tree, grass & weed pollen in an urban area (Montréal) using flow cytometry data & a machine learning algorithm (specifically random forest). They identified granularity & fluorescence parameters from violet (Violet610_A) & blue (PB450_A) lasers to be the most distinctive features for discerning a pollen species, though species accuracy (F1 = 0.76) in the random forest-model was lower than genus accuracy (F1 = 0.90). I generally found the manuscript to be well written, with few spelling, grammatical issues. This study presents a strong step toward scalable pollen classification using non-imaging cytometry. However, some concerns and requests for clarification are pointed out below:

Major comments:

Line 94-95: For this study, I understand that the purity of the pollen can be critical in analyzing the data obtained from the flow cytometer because the proportion of non-pollen particles that can be similar in size or fluorescence intensity can influence the classification. However, there is no data to validate the pollen's purity using other methods. For example, would it be possible to compare the pollen purity by an image analysis of the extracted pollen subsample under a light/fluorescence microscope?

Line 143-144 = Since the data is unbalanced, the authors chose to use synthetic minority over-sampling to normalise the data. However, I was under the impression that oversampling can cause model overfitting & an inaccurate representation of the smallest minority classes (~300 – 35000 is very unbalanced). Were steps taken to avoid/ensure that didn’t happen? It may be at least worth a mention in the discussion.

Line 145-146 = Reads as synthetic minority over-sampling is done on the samples, then the dataset is split into 70% training & 30% validation. If so, would this not lead to overfitting & inflated precision metrics, since the training data is used in the validation set? Perhaps I’ve misinterpreted what’s written, or this doesn’t matter. If so, clarification may be needed in the order of steps taken.

Line 160 = While F₁-scores are useful for measuring model performance in a single metric, the likely strong class imbalance and use of oversampling would suggest that including metrics such as PR AUC in addition to F₁-scores would provide a more complete summary of the model's performance (Saito & Rehmsmeier, 2015 https://doi.org/10.1371/journal.pone.0118432).

Minor comments:

Lines 103-106: The authors state that the pollen's fluorescence depends on the fluorescent proteins on its surface (my understanding is that these are not proteins). If so, please provide a reference.

Line 148 = At the beginning of 2.4. It is stated that four supervised classification algorithms were tested, and the random forest performed best. There are many types of random forest classifiers, such as random forest by randomisation, which deals well with unbalanced/noisy data. Which random forest classifiers (Breiman?) were tested? It may be worth mentioning why the particular random forest was chosen over other random forest classifiers.

General comment

This preprint describes an innovative method for measuring and identifying pollen grains using a standard laboratory flow cytometer and machine learning tools. The authors made considerable efforts to collect and prepare a comprehensive pollen reference database, which they then used to test their method. The results are promising, showing that liquid flow cytometry coupled with machine learning algorithms could become an effective alternative to manual pollen counting and identification. I only have a few questions related to the content and some technical corrections. In my opinion, no major changes are needed.

Specific comments

• l.88-91 “Their inclusion enabled the model to learn to discriminate tree pollen from other common airborne particle types, as real-world environmental samples typically comprise a heterogeneous mix of tree, grass, and weed pollen, along with various non-pollen particulates.” Are non-pollen particles a majority in your samples? What proportion did you observe?
• l.104-106 “Depending on their peptide composition, these proteins absorb light at a certain wavelength and emit light radiation at a different wavelength in return producing a characteristic fluorescence signature that varies among species.” Is the fluorescence signal expected to differ from airflow cytometry due to the suspension in PBS?
• l.106-108 “For each laser, avalanche photodiode (APD) detectors measure the intensity of light emitted at different wavelengths using ten filters: 450/45, 525/40, 610/20 (violet laser), 525/40, 585/42, 690/50, 780/60 (blue laser), 660/10, 712/25, 780/60 nm (red laser).” It would be nice to have an explanation on these values, e.g., for 450/45 what does 450 and 45 represent? Most people in the aerobiology community are not familiar with these measurements.
• l.108-109 “In addition to fluorescence, two scatter parameters were recorded to describe particle morphology: grain size and granularity.” Give a precise definition of what size and granularity represent in this study as they can be misinterpreted. Is the size value the maximum axis of the pollen grain? What values do we expect for granularity?
• l.125-126 “This resulted in a total of 25 parameter values per particle.” Since the previous sentence is quite long I would summarise here the number of parameters: “This resulted in three values for size, two for granularity and 20 for fluorescence, with a total of 25 parameter values per particle.”
• l.127-129 “For each species, pollen grains were manually separated from debris using scatter density plots (size vs. granularity) and histograms of all fluorescence features.” This is a very interesting approach that needs to be shared in the aerobiology community!
• l.140-141 “Among these, the Random Forest algorithm showed the best performance and was therefore selected for subsequent analysis.” Which criterion was used to determine that Random Forest showed the best performance? F1 score? Please state it briefly in the text again.
• l.160-173 It is very interesting to compare the species level with the genus level. In these two paragraphs, you describe the taxa that performed with reduced accuracy. What about those that performed best?
• l.212-214 “The models achieved high classification performance (F1=0.76 at the species level and 0.90 at the genus level) highlighting the potential of this approach as a scalable alternative to traditional microscopy for pollen identification.” For clarity, it would be good to state that this performance was achieved under lab conditions.

Technical corrections

Technical corrections are highlighted in the attached PDF file.