Review comments on “Parallel use of threshold parameter variation for tropical cyclone tracking” by Enz et al.

Manuscript ID: gmd-2022-279

Recommendation: Accept with minor revisions

General comments:

This manuscript investigated using multiple combinations of different threshold parameters for tracking tropical cyclones from the 20 member ensemble regional ICON-LAM model simulations for the 2013 North Atlantic (NATL) hurricane season. It is shown that using multiple combinations of thresholds is beneficial in tracking/identifying TC genesis early stages as well as the extratropical transition or decaying phases. Overall, the manuscript is well organized and prepared. I only have a few minor concerns (see details in the specific comments below) before the manuscript can be accepted for publication.



Specific comments:

1. The tracker input data used in this work were from the 20 member ensemble ICON-LAM model simulation for the 2013 NATL hurricane basin/season driven by ERA5 reanalysis data. In the manuscript, it seems to me that, the authors only showed/discussed the tracking results of these 20 member ensemble ICON-LAM simulations, but did not compare the tracking results against the observed (best track) data (or even just the tracking results by using the ERA5 reanalysis data though with relatively low horizontal resolution). I would suggest adding some analyses/comparisons to demonstrate the capability and effectiveness of the ICON-LAM model simulation together with the tracking algorithm used here in terms of capturing/reproducing the TCs for the 2013 NATL hurricane season.
2. Could the authors comment on how the different tracking threshold parameter values are determined? Are they specifically chosen for the ICON-LAM regional model configurations (e.g., 13-km grid spacing, etc.)? In other words, are these threshold values generic to be used for other model configurations/simulations?
3. When analyzing/comparing the tracked ACEs from the ICON-LAM ensemble member simulations, I would suggest adding the comparison against the ACEs derived from the best track data (e.g., HURDAT2) as well.
4. I know the focus of this work is on storm track, intensity, and development phase (weak, strong, genesis, decaying, transition, etc.) tracking. I was wondering if the tracker used in this study can also track storm structure and size metrics (radius of maximum wind, radii of 34-kt, 50-kt, 64-kt winds, etc.).

This paper applies an approach of running the same tropical cyclone (TC) tracker with multiple combinations of threshold parameters across a multi-member regional ICON ensemble. The data is then "merged" allowing tracks to have an "identification percentage" -- that is, the number of combinations that flagged the particular storm as a TC. Storms are manually classified into different genesis and termination categories, with panel plot examples shown of each. The authors described potential false positives and also explore changes in accumulated metrics (e.g., accumulated cyclone energy, ACE) based on whether these false alarms (or weaker storms) are included or the allowable translation speed is changed. They generally find integrated metrics are relatively insensitive to these issues.

The idea of looking at "matched" detections across the parameter space is interesting and the work is suited to GMD since this pertains to the development of an algorithm to assess climate model performance. That said, there are many areas where I think the manuscript should be improved both scientifically and from a readability standpoint before it can be considered for publication. In particular, the model setup and data generation are poorly described and it is not 100% clear exactly what members are being analyzed and when. The work is also under-referenced, with previous papers exploring aspects such as parametric sensitivity and cyclone termination (extratropical transition) that are not cited or discussed in context with the results here. Improving both would heighten the paper's potential impact and make it more applicable for developers of numerical models and TC researchers. More comments are below.

Major comments:

How the simulations are performed and analyzed is quite unclear. Briefly, this is how I am interpreting these results. 20 ensemble members of the summer season 2013 are generated, with each ensemble member running identical LBCs (and surface forcing), except with the initial conditions shifted in time. Each ensemble member then spawns its own TCs internally which are then tracked -- that is, the simulations do not purport to be a reanalysis of the 2013 time period. Each combination of tracking parameters is used to track the TCs in each ensemble member. Only a single member is then analyzed (i.e., none of the figures seem to be showing a composite mean, but rather a snapshot from a single run) *except* for Figures 11 and 13, which include information about all 20 members. That is, the "identification percentage" reported in early figures is how many of the parameter combinations matched for a single TC snapshot in a single ensemble member.

The initial times are listed as having occurred during May 2013 but line 102 mentions June-November. Do the authors eliminate the beginning of the simulations (i.e., first 1-4 weeks) for spinup and start analysis on June 1 for all runs?

The simulations are run with ERA5 boundary conditions, do we expect TCs to be matched *across* the ensemble members or just the ACE to be correlated as in Fig. 11? Put another way, do we expect Fig. 2 to look similar if we use a different ensemble member, or are the LBCs too far away to influence TC genesis and tracks within the center of the domain? To me, these simulations seem like a precursor to a seasonal prediction system (i.e., force with some conditions but acknowledging the model can generate different realizations of TC activity underneath these forcings) but I could be wrong.

Regarding parametric sensitivity, both Horn et al., 2014 (Tracking scheme dependence of simulated tropical cyclone response to idealized climate simulations) and Zarzycki and Ullrich, 2017 (Assessing sensitivities in algorithmic detection of tropical cyclones in climate data) are under-referenced. Both also looked at how changes in tracker settings could impact tracked TC statistics. It would be beneficial to link some of their findings regarding parameter sensitivity to those here. For example, the former paper found that threshold differences were the most important contributor to differences between modeling centers with different algorithms and the latter paper also found that integrated metrics (such as ACE) were relatively insensitive to tracker configuration. Ullrich and Zarzycki, 2017 (TempestExtremes: a framework for scale-insensitive pointwise feature tracking on unstructured grids) contains a review of previously published trackers and associated literature that may be worth exploring (e.g., adding additional references in the introduction and comparing/contrasting top-level findings in the results and discussion).

Section 5. There is work in this space regarding using cyclone phase space to help determine genesis/lysis. Two papers that come to mind are Bieli et al., 2020 (Application of the Cyclone Phase Space to Extratropical Transition in a Global Climate Model) and Bourdin et al., 2022 (Intercomparison of Four Tropical Cyclones Detection Algorithms on ERA5). The latter is relevant to other aspects of this work (they utilize a hit rate/false alarm approach against pointwise observations to understand parametric and dataset uncertainty).

Other comments:

Fig. 3, why does the identification percentage go down (red, bottom) when the storm intensity is going up (L to R, black curve up, blue curve down)? Are the authors discussing this when they mention "The allowance for this displacement is sensitive to the warm core threshold parameters, which is reflected in the reduction of the identification percentage for this time step. This in turn shows that the identification percentage is not sensitive to TC intensity alone"?

Why are the identification percentage lines not shown at the bottom of Figs. 6 and 7 as with the previous figures? I would assume these would decrease to the right with time?

The idea of "parallel" in the title needs to be reworded. Most people seeing that are going to interpret this as code that has been parallelized (using MPI, for example) when I believe the authors imply they are running the same code with multiple parameter combinations and then combining the results into a single track dataset.

It wouldn't hurt to add further clarity to the section describing the algorithm. For example, it appears all local minima satisfying criteria 1 are first found. Then the columns above them are scanned for vorticity. Assuming both of those checks are satisfied, the column is again checked for a T_anom maximum. All these storms are considered potential storms and then "glued" together as a second step dependent on the tau (duration) threshold.

There has been some recent work that shows that sea level pressure is better simulated in atmospheric models (Roberts et al., 2020 "Impact of Model Resolution on Tropical Cyclone Simulation Using the HighResMIP-PRIMAVERA Multimodel Ensemble") and is a better correlate to damages (Klotzbach et al., 2020 "Surface Pressure a More Skillful Predictor of Normalized Hurricane Damage than Maximum Sustained Wind"). It may be interesting to categorize storms by PS which would also eliminate the below issues regarding maximum wind and scanning radius. 

Lines 20-21 can be rephrased since the authors argue that manual tracking is complicated by subjectivity but then they undertake a manual tracking in part of the manuscript.

Line 99. Describe R03B07 -- I assume ICON is on an unstructured grid, hence the need for remapping to a regular grid?

Line 108. Do the authors mean geopotential height? I am not sure why surface geopotential needs to be prescribed at the lateral boundary as this is commonly a constant surface boundary.

Line 190. See Stern and Nolan (2012) "On the Height of the Warm Core in Tropical Cyclones" which would provide more context than the large range (most of the troposphere) described here.

Line 204. The 113 TCs are detected across the 20 ensemble members? So about 5-6 TCs per member?

Line 255. Why is 100km chosen? In observations, TCs are generally O(1000km) wide and most tracking algorithms look out at radii from 200-500 km. 100km can be inside the RMW of even mature TCs (see annular storms). I understand the concern with picking up non-TC wind speeds, but that issue would seem to be problematic with extremely large radii, not O(250-500km).

Line 329-331. How was this determined? If this was a manual process (as I assume it was), what factors were taken into determining a false positive? How were "edge" cases (storms that were perhaps subtropical in nature) handled, or were all the false positives as obvious as this one?

Line 354-355. This would seem to be a solvable problem in that local minima could be merged or an offset between sea level pressure minimum and vorticity maximum can be allowed.

Line 426-430. See above comments re: r = 100km for vmax calculation.

Line 445-446. See the discussion of Bieli and Bourdin (and references therein, I assume) above, since this approach has been applied previously in tracking algorithms.

Figures. The labels are very small and hard to read. Also, the panel titles can be cleaned up (ex: 08_ref 26 can be eliminated).

Line 414. Needs to be \citep{}.