This manuscript provides a novel and necessary technique to improve the finding and targeting of tephra layers (and other particle types) in ice core samples. The authors did a rigorous analysis by imaging synthetic and natural particles and used machine learning to identify the different particle groups. The size of the particles that they can image is greater than most other grain size analysis techniques being used today in ice core research.  Marine particles (diatoms and sponge spicules) are a little underrepresented but can sometimes be larger than the tubbing diameter (80 microns). Complex rock fragments (i.e. lava flows) are also not really identified in this study. Many particulate layers in Antarctica are wind-blow rock fragments and would have a distinctive shape when compared to mineral fragments and glass shards. This along with adding more types of pollen to their image dataset could be an area for future improvement.

There are a number of comments in the attached pdf that need some clarification or elaboration. Many of my comments deal with the FlowCam setup. It would be really helpful if there was a picture or diagram of the FlowCam setup, even if in the appendix. I was confused by the orientation of the tubbing and the gravitational settling of larger particles. A cross-section diagram of the tube would help explain the imaging volume (41.8%) and the problem with large and blurry particles.

Being able to image particles assess their grain size and give them a particle type (e.g. tephra, dust, etc.) before a tephra specialist gets the samples is extremely helpful and will improve the number of tephra found in ice cores and will decrease the time needed to find said tephra.  However, the authors do not discuss how to physically capture the particles after FlowCam analysis. Capturing these particles so that they can be analyzed by SEM or EMPA is the most important part of this type of work. It is great to know the particle type and grain-size distribution but this method falls short if geochemistry on the particles is not obtained. It would be great if the authors would elaborate on capturing these particles. Their goal is to help both the CFA and the tephra communities. The CFA community doesn’t like to run particles through their MS and the tephra community wants those particles. This method can be extremely helpful in spotting these interesting intervals.

Overall this is an excellent paper that addresses a need in the ice core community. With some minor corrections and a few elaborations, this manuscript is ready to publish. I hope to see this type of FlowCam analysis being used in more labs and on more cores.

SUMMARY

In this work, the authors present a framework to detect and classify ice core particles via a proprietary solution, i.e. a FlowCam instrument (Flow Imaging Microscope), combined with Deep Learning. The authors used the FlowCam to create the image data set as well as a set of numerical features, created by the FlowCam software. The data set is comprised of 7 classes, including dust, two tephra classes, three pollen species, and basically an extra class for contaminations and blurry objects. The model is hybrid; it consists of a MLP and a ResNet-18, which functions as the actual feature extractor, pretrained on ImageNet. The model is fine-tuned by training all layers. The numerical features are fed into the MLP, which are ultimately concatenated with the output of the ResNet-18. The authors achieved an accuracy of 96.8% across all classes on the test set.

PROS

The biological and geological background is well written and established. The general writing quality is exceptional. The need for an automated solution to the labor- and time-intensive manual process becomes evident as well as the benefits of such a system. The overall story is convincing and probably a useful benefit in the scope of its discipline.

CONS

Despite minor writing mistakes, I must address a couple of issues with the paper:

1. A ResNet-18 is selected as the main feature extractor; why is that the case? Are smaller nets better than larger, are e.g. Residual Blocks superior to Dense Blocks (DenseNet)? What made the authors chose this architecture?
2. Are the authors dealing with an object classification or detection problem? There is a distinct difference between the two: The first aims at classifying images of isolated particles, while the latter aims at detecting (counting) and classifying a number of particles in an image. I assume it is the first (also due to the images shown in Appendix C), then how are the segmented/isolated particle images generated? By-hand, by the FlowCam software, or a different method?
3. The authors mention in line 138 “false positives”. This hints at a typical metric from object detection tasks, typically depicted via a confusion matrix. However, this also includes more important metrics than accuracy (such as F1, precision, etc.). This is confusing (see also point no. 2).
4. Why are the images downscaled to 128x128 pixels and from what original size? Pretrained nets (including ResNet-18) on ImageNet usually take inputs of the size 224x224.
5. The numerical features are based on geometrical properties and the product of a black-box by the FlowCam software. Are they necessary? Deep Learning methods operate extracting intrinsic features (if enough data is given) “by themselves”, feeding the network additional “hand-crafted” features is contrary to the very idea of DL. Is there an advantage in using the numerical features? (e.g. is there proof that they increase the accuracy?)
6. The work lacks some reference to practical computer vision applications in related fields. I would recommend including at least “Pattern recognition methodologies for pollen grain image classification: a survey” (Viertel, König, MVA Journal, 2022). This would also bring more insight into the problem mentioned in point no. 5.

CONCLUSION

If the points are addressed in the manuscript and the authors argue convincingly for their design choices, I recommend it for publication.