This manuscript present a new algorithm to detect seismic events. The method is based on the classical sta/lta detection method. 

It runs the algorithm using a wide range of time windows (tSTA and tLTA) to build a hybrid detection function. 

This method allows to detect a wide variety of seismic signals covering a huge range of signal durations (0.1 - 10000 s).

The method is applied to a dataset from the Whillans Ice Stream. However, this manuscript does not provide new information on the source of these seismic signals. 

Most possible "stick-slip" events are already known (Pratt et al, 2014) and their correlation with tides has already been extensively discussed.

I am rather disappointed by this paper. 

First, it is purely methodological with no result on glacial processes. 

The method could also be applied in many fields of seismology (landslides, volcanoes, ..) that also exhibit a large variety of seismic signals.

It could thus better fit in a journal on seismology.

Second, I am not fully convinced by the advantages of the method compared to other methods.

The main advantage is to automatize what many researchers do by trial and errors. 

Many researchers adapt the time windows of the STA/LTA methods (Short and Long Time Average of seismic energy) in order to detect most events 

that can be observed by eye when looking at seismograms and spectrograms, while simultanouesly decreasing the rate of "false detections", 

ie, anthropogenic and environnemental noise, teleseisms or other types of events different from what they look for.

But I think that this first step of looking at seismic data (on a small subset of the dataset) is essential to discover different types of events and to select signal of interest.

Also, the method only considers two parameters (tSTA and tLTA) but does not discuss the other parameters: the minimum ratio of short and long-term energy used to define an event and the frequency band.

I guess the authors do not filter the data, while it could be an efficient way to remove noise events and to detect weaker events by selecting the frequency of interest and where the signal/noise ratio is largest.

I feel that the method allows to detect more events but most of these events are maybe noise, such increasing the work of event classification, which is the most difficult task.

I believe that a simpler STA/LTA method with well chosen parameter could be almost as efficient than the proposed multi-STA/LTA algorithm, while reducing the number of "false detections".

At least, the authors need to demonstrate that their method detects more events but without increasing the fraction of false detections.

The multi-STA/LTA algorithm is compared with the standard STA/LTA method, but only for the two extreme models (very short or very long time windows, l232).

Why not using all models or the average model?

The manuscript is often hard to read and understand, many points should be clarified (see minor points below).

The manuscript stops when things could start being really interesting.

What are the newly discovered "stick-slip" events? Why were they not detected before? Are they weaker than the others or do they have a different waveform? Could you try to locate these new events?

What are the "tremor-like" signals mentionnent on l326?

Details and minor points:

Figure 1. Plots (a) and (b) could be removed, all information is also on the other subplots.

Algoritm description, section 2.1. 

The classical recursive STA/LTA algorithm should be described (even it is described in the cited references) as it is the base of the multiSTA/LTA method.

l118. What are the values of the minimum STA/LTA thresholds (trigger and detrigger) used to define an event? 

How are they chosen? Why not optimizing these parameters as done for the time windows?

Fig 1 suggests the threshold is fixed at 3 and is the same to trigger an detrigger an event. Did you try other values?

I don't understand eq. (1) and point (3) (l116). Is the hybrid function the average (as in eq(1)) or the maximum value (l116) of all single-parameters STA/LTA functions?

l133. I don't understand what represents epsilon?

Section 3.2 should be moved to the "method" section 2. It describes how the catalogue is compiled and is not specific to the Whillans Ice Stream catalog.

l194: I don't understand this sentence: " The reference time precedes this arrival time by half the network time for the N closest seismometers."

What is the "network time"? What is the value of N?

l195 "We also take into account that multi-STA/LTA will decompose events into smaller events based on amplitude variations."

I think this is a drawback of the method; and a simpler STA/LTA method with well chosen parameters and with a detrigger threshold lower than the trigger threshold may avoid this problem.

You should describe in more details how you merge "overlapping events" into one event to obtain the correct event duration.

l213 "Taking uncertainties in start time into account, we label 140 events as stick-slip. Four of those events are determined as additional to the Pratt et al. (2014) catalogue from a manual reconnaissance."

How do you know that these 4 events are "stick-slip"? Could you show examples of a newly detected "stick-slip" event and a known stick-slip event? 

l219 "What means "a.u.": arbitrary unit? 

l267 "The distribution of the peak amplitude occurrences provides source mechanism information". 

Could you specify which "source mechanism information"? Do you mean the Gutenberg-Richer b-value?

l268 "However, in cryoseismology, the actual magnitude cannot be determined because the material strength, slip distance, and area of slipped fault are usually indiscernible."

I don't think this is a big issue and a major source of uncertainty on magnitudes. The problem is even worse for earthquakes that occur at depth were material properties are less constrained.

l271 "The maximum of the occurrences ". Do you mean the maximum of the distribution of event amplitudes?

There are several similar sentences where I was not sure to understand what the authors mean.

l287 "As a drawback to this approach, a small number of event groups might be catalogued under a single energetic reference event even though the source mechanisms could be different."

Yes indeed, many processes produce a wide range of signal peak amplitudes (Gutenberg-Richter law). 

I think that the event duration or frequency content is more useful to identify the source process.

l289 "It is possible that events in other locations of interest for cryoseismology have event types with substantially different seismic signatures than those of the WIS (on which our simulated waveform population was based). "

Yes, indeed. For instance, basal stick-slip events have duration ranging between 0.1s and 1000 s and frequencies between 0.01 Hz and 1000 Hz (see Podolskiy and Walter 2016 for a review).

Fig 6 "The multi-STA/LTA algorithm combines advantages of the other algorithms, as it is able to match, and improve upon, the detections achieved by RECmin"

But are all detections real events or is there a significant fraction of false detections (noise)?

Fig S3: The signal with a frequency of 0.01 Hz is consistent with stick-slip, but it could also be a teleseism

Fig S5: Why filtering the signal? At least the spectrogram should be shown for the raw unfiltered signal.

l303 "we have manually identified events of stick-slip origin": could you explain how you distinguished "stick-slip" events from other types of events?

l306 "The general trend between peak amplitude and duration (7, top) and energy and duration (7, bottom) of events is consistent with the positive linear association expected from cryogenic sources "

This is very general and true for many different source types. So it is not useful for classifying events. 

l325 "The events of lower energies,..., that occur for long durations (bottom) suggest the presence of harmonic tremors in the catalogue."

This could be interesting! Could you show an exemple of seismic signal and spectrogram ?  What could be the source (slow-slip event, water flow, storm ...)?

Section 4.2.1. The correlation with tides and temperature is interesting and is a good way to investigate the source mechanisms. 

But it should be done after classifying events in different types, ie, removing known stick-slip and teleseisms.

Rather than amplitude, I think that frequency content  (average frequency and width of the spectrum) and signal duration could be better parameters to discriminate source mechanisms.

Fig 7. You could show only plot (a) as it contains almost all information shown by the other plots.

l361: "Further, the production of near-comprehensive, reproducible event catalogues is a critical step towards standardized glacier monitoring as comparative studies between locations are enabled."

I agree that comprehensive and reproducible catalogs are valuable, but I think that standard methods (simple STA/LTA or template matching) can already produce such catalogs.

I think that each glacier is different and that all algorithms and parameters need to be adjusted for each case study. 

I also think that using different methods may allow to detect events that are still unknown.

The main problem is not the detection of events, that can be easily automated and reproduced by others. 

The classification of events is much more tricky and subjective, often done "by eye" without objective criteria. 

I understand that this is the goal of your companion paper (l373) and I am very interested to see how a fully automatic machine learning method can perform.

l370 "The new catalogue will find utility in guiding conventional glacier seismology." Can you explain how? 

L408 "We find a partial association of seismicity with the tidal cycle,". This is not a surprise since the catalog contain many already identified stick-slip events that are known to be driven by tides.

L12, L409 " We find a slight association with ice surface temperature, as an indicative example of one atmospheric observable.". 

I don't see such a correlation when looking at Fig 9. This "association" should be quantified and tested using a statistical test.

L414 "semi-automated approach". When reaching the conclusion I still don't clearly understand which part of the detection method is not automatic?

The authors present an exciting study that describes a “recognisance” algorithm for detecting seismicity from a range of seismic signals. The method is novel and the concept of a recognisance algorithm that is sensitive to a range of seismic signals will likely be of much use to the community.

However, although I would like to see this work published, I think the wording of the manuscript needs some work before it is ready to accept for publication. Firstly, the novelty of the algorithm needs to be clarified. This is probably a minor point, but at the moment the method appears to already be published (Turner et al., 2021), but the authors do not make this explicitly clear throughout the paper. I’d like to see the original methods paper properly acknowledged where the method is first introduced, and the tone of the paper changed to reflect that the method is applied to cryoseismology here rather than introduced as a new method. Secondly, I remain to be convinced by the concept that such a deliberately broad method can outperform a more specific method suited to one task (e.g. basal icequake detection). Perhaps it can, but I see no evidence in the paper to back up the claim made in the conclusions that the method presented can provide a “near-comprehensive event catalogue”.

To conclude, I think this work will be a valuable contribution to the field and I don’t find any issues with the results themselves. However, the text needs to be somewhat revised to tone down the claims made. Assuming the authors are happy to do this, then I would be very happy to see this paper  published.

Comments:

The introduction does seem to be written in rather a bold way (e.g. “Since STA/LTA and correlation-type algorithms have enjoyed only limited success when applied to environmental seismology”). Both STA-LTA and cross-correlation methods are effective when used in certain ways, as the authors elude to and indeed that is the premise of this paper. I’d suggest toning down the limitations of these algorithms a little, since actually there has been much success in using these algorithms carefully, as part of broader methods. Indeed, I cannot think of any passive seismology method (other than manual detection) that is not at least somewhat built upon one of these two foundations. I think it would be useful to outline the scope of the study from the very beginning of the introduction, to clearly clarify the scope of the work to the reader before making the perhaps bolder claims.

From the introductory text, I was expecting Section 2.2 to be its own section. However, I can see the logic behind having it as a subsection of the overall algorithm method. Regardless of the structure, I think it would be useful to provide more details on the analysis of the algorithm testing results presented in Fig. 2,3. There is a lot of information held in those two figures, but they are not yet adequately described in the text. In other words, I was excited to read that the algorithm would first be tested on synthetic data, but was then left a little disappointed that the key findings are buried in the figures without much explanation of what they show.

L183-187: Perhaps it is common to only use the vertical component. However, best practice for any body wave data should be to use both vertical and horizontal components. However, mixing the three components via the Eucllidean norm might cause one to loose information about whether a phase is a P or S phase. While information loss is not inherently a problem for detecting events, incorrectly identifying S-wave phase arrivals as P-waves would result in false detections, since they are likely from the same event. I’m not suggesting the authors should revisit this component of their method, but I think they should be clear about the limitations, especially the possibility of false triggering. I’d suggest it would be beneficial to also discuss somewhere how their method might be developed further to act on the vertical and horizontal phases separately. If the authors would like any pointers to literature describing how to use the vertical and horizontal component information for P and S wave association in detail, including how a firn layer can affect such results, then this paper and references therein provide further details (https://doi.org/10.5194/egusphere-2023-657 ). (No requirement to cite this work, just a potentially useful paper that covers the point raised).

L189-190: Some clarification on the novelty of the method is required. Until this point in the paper, I was under the impression that the authors were presenting a new algorithm for event detection. However, from a glance at Turner et al. (2021), it looks like the value of this work is more in showing how the cited algorithm is implemented? Could this be clarified, and if the implementation is originally from Turner et al. (2021), then that paper be clearly referred to in Section 2.

L200-202: Perhaps too technical a question, but more for my interest: Are the signals instrument-response corrected before being passed through the algorithm? If not, then “energy” might be better referred to as “an approximation of the energy”, since different frequencies might exhibit different amplitude responses dictated by the instrument transfer function (response). Normally I wouldn’t raise this point, but since the authors are attempting to describe as broad an event detection algorithm as possible, frequency response could become important in certain instances. Maybe at least worth making readers aware of this point.

Figure 5: It would be nice to see more detail in each of the time-series. I cannot dicern any differences from the plots in this figure. Maybe for each, the authors could include an inset figure zooming in on any differences in the first arrival, perhaps plotting all three signals over one another? Otherwise I’m left questioning what difference the multi-STA/LTA algorithm makes. In summary, I imagine it is better, but the figure does not currently communicate that.

L276-277: Great aim. However, this sentence is very much left hanging. I’d like to see some text introducing how the authors will justify how the work presented here meets that aim in the remaining discussion.

Section 4.1: The biggest limitation of this work to meet its aim is that such a broad search algorithm does likely not perform as well for certain event detection scenarios, compared to more specific methods. For example, a coalescence/stacking based migration algorithm will, by design, be better for filtering out false triggers caused by noise. Therefore, taking basal icequakes as an example, the authors method is unlikely to outperform a migration method. This does not detract from the method presented by the authors, since a general recognisance method is definitely very useful in many contexts. However, it is definitely a limitation that needs to be mentioned, because it really does present a barrier for “a consistent approach to the generation … of event catalogues” (L276-277).

L411-412: This statement is definitely too bold. I am not convinced that one can ever develop algorithms that produce “near-comprehensive” event catalogues. However, more specifically, a general recognisance algorithm, which this is presented as, is unlikely to outperform a specific algorithm for a specific purpose. The algorithm presented here is definitely a valuable contribution to the field, but I see no evidence in the manuscript that it can produce “near-comprehensive” catalogues of seismicity.

Minor comments:

There should be spaces between units (e.g ma^-1 should be m a^-1, otherwise it is technically milli years).

L19: “exceptional” – rather emotive language. Consider removing.

L23-24: I think the definitions of the two event detection types are a little narrow. I’d view STA/LTA algorithms as just one subclass of any algorithm that searches for a peak in energy with particular frequency content (related to the length of the STA window). These can then be used to detect phase-arrivals, or be used in combination with more sophisticated phase associators, or coalescence-based algorithms to improve detection. Perhaps worth somehow very briefly mentioning this (as an STA/LTA algorithm on its own is not very useful/causes lots of false triggers).

L23-24: Further to the point above, one could consider array-based methods as a common earthquake detection type too? Not sure if this is worth mentioning, but picking phase arrivals on individual channels is only one class of earthquake detection. If the authors do decide to mention this, then examples of some relevant papers are included in the introduction to this paper: https://doi.org/10.5194/egusphere-2023-657. (sorry to refer to this paper again – just easiest way to point to the relevant literature cited within it).

L43-44: I’d say that cross-correlation algorithms are inherently also similarly prone to missed detections since they are typically based on the assumption that similar events occur within a catalogue.

L72: Is QuakeMigrate a spectral-based method?

L122: “event” – would it be perhaps better to refer to it as an “event phase arrival”?

Figure 1: Isnt the data in Figure 1b also plotted on Figure 1c? Not sure this section is therefore required. Similarly, Figure 1a could also be removed as all the information is contained in (d).

L170: The minimum distance rather than the maximum distance is probably the more relevant number.

L199: That reference is not to software documentation, but to a paper describing the software. Pointing to any software should really be done via a software repository DOI in the Acknowledgements. However, I think the authors are actually refering to the paper here, and so should simply remove the word “documentation”.

L236-237: I think this sentence doesn’t make sense with the word “event” at the end. Apologies if I misread, but would it be possible to reword if indeed it doesn’t make sense?

Figure 5: Add subplot labels, then refer to accordingly in the text (rather than left, right etc). That way the UTC time stamps can be remove from the text. Also, the caption is too long. Consider shortening.