(see attached pdf)

Review of egushpere-2023-2003:

Implementing a dynamic representation of fire and harvest including subgrid-scale heterogeneity in the tile-based land surface model CLASSIC v1.45

Summary

The authors present a subgrid tiling method for a land surface model to improve simulations that include fire and harvest disturbances. Each grid cell is divided into tiles based on two potential disturbances per year. Each tile has a representative set of pfts (including veg height) and a time since disturbance. The authors show that multiple tiles (as opposed to a ‘single’ tile) can have a dramatic impact on biogeochemical outputs, and that increasing the maximum number of tiles eventually reaches an asymptote with respect to changes in outputs. The authors further explore dynamic tiling parameters to find an optimal configuration. They conclude that this is a viable approach for land surface models to reasonably capture more details associated with subgrid vegetation disturbance processes.

Overall response

This is a good example of model advancement that increases detail and complexity to achieve greater accuracy without requiring major restructuring of input data. It also demonstrates how much of a difference it can make to try to more accurately represent vegetation change at a subgrid level. It does require some clarification, I am not convinced by the choice of optimal setup, and I it is unclear whether this method actually increases the accuracy of the model. My main concerns are outlined here, with more detailed comments following.

1) Some of the text is unclear, particularly in the methods section. See details below.

2) The optimal configuration is selected simply for computational efficiency, rather than taking into account the potential model response. But the response analysis shows that the responses can be quite different, while the computational efficiency appears nearly the same for all 12-tile configurations. If the reference configuration is truly believed to be a more accurate representation of the processes, then it should factor more strongly into this decision. In particular, the two disturbance outputs have a very poor response with the chosen optimal configuration, in relation to the reference. One challenge here is that there is no accuracy or skill assessment, so selecting an optimal configuration is lacking the dimension of model accuracy (see next point). Another is the units of the computational efficiency: the reader cannot tell whether a one second difference per cell actually matters. Doe this difference mean the model takes either 15 or 18 hours to run 30 model years, or 5 or 18 hours to run 30 model years? If it is the former, then you want the more accurate configuration. If it is the latter then you have to consider resource tradeoffs.

3) Does this structural advancement improve model accuracy? The assumption is that by representing finer resolution disturbance the accuracy of the simulation should improve. But the one comparison with above ground tree biomass does not indicate any model improvement with this structural change, but it does take more computational resources. So how do you justify the increased complexity? Key outputs are clearly affected by this approach, but do you want to use this approach if it reduces model skill? This may or may not be required in the context of GMD, but I suggest running your outputs through some sort of benchmark or skill assessment to show that this approach is a worthwhile advancement.

Specific suggestions/comments

Abstract

line 30

but you don’t show what the model biases are or whether they are reduced. so is there improvement?

Introduction

line 120:

what processes are you referring to? all of the one mentioned in this paragraph (ranging from disturbance to energy flux to model algorithms)? additional ones previously mentioned (e.g, vegetation productivity). the most recent processes mentioned are tile creation and merging.

Methods

line 125:

“Our study domain encompasses all of Canada south of 76N”

line 130:

“In Canada, annual, contiguous timber harvest events remove 98+-…”

line 132: see line 130

line 135:

How does harvest account for only 0.2% of stand replacing disturbance if 52% of the forest is managed?

line 141:

Be clear that CLASSIC couples CLASS and CTEM; according to the next paragraph, CLASSIC isn’t merely based on them.

line 156:

does “canopy-covered ground” mean that it does canopy energy exchange, or is this done by CTEM?

lines 175-182:

several other subgrid papers exist. here are a couple of examples.

subgrid and surface energy balance:

Hao et al 2022. Impacts of Sub-Grid Topographic Representations on Surface Energy Balance and Boundary Conditions in the E3SM Land Model: A Case Study in Sierra Nevada. james, 14(4):e2021MS002862. https://doi.org/10.1029/2021MS002862

subgrid and water, fluxes, energy balance:

singh et al 2015. Toward hyper-resolution land-surface modeling: The effects of fine-scale topography and soil texture on CLM4.0 simulations over the Southwestern U.S. water resources research 51(4):2648-2667. 

https://doi.org/10.1002/2014WR015686

lines 206-207:

While it is technically necessary for the new tile to  not exceed the available space, this limit does not make sense in this context because it would require all of the candidate tiles to be merged to reach this limit. There is a semantic challenge here where “splitting” multiple tiles also requires merging the split-off areas. You may consider tile “creation” and “joining.”

line 233:

eq 4 appears to be equal to a unitless 1. I think you need to remove the t term from the denominator.

lines 234-253:

The description and variables do not match the equations, which makes this section confusing.

It also is not clear that there is only one parameter for allowing merging, and then a second one for preventing merging. When and how are the rht and tpp set? is tpp a minimum number of total tiles to keep? Does tpp simply retain the shortest veg tiles in order to meet this minimum number? is rht the threshold, or is there additional calculation required to get the threshold (and how is it calculated)? what happens if rht and/or tpp are not set?

lines 296-392:

you may want to reiterate that these fractions are specific to canadian forest harvest and processing.

lines 313-320:

if this is static land cover, what year(s) is it based on?

line 318:

prescribed land cover can vary over time; this is static land cover

lines 358-366:

this is unclear and confusing. figure 3 helps somewhat.

lines 370-375:

this information should probably be up where rht and tpp are introduced. it should help clarify what rht and tpp are, what they mean, and how they are used. here rht and tpp sound very different from how they are introduced. also, how is preemptive merging different from regular merging. the previous section describes only preemptive merging.

line 382:

this is related to lines 322-366. i appreciate that you do your best to develop fire and harvest drivers for the entire simulation, but you end up with three very different disturbance regimes for each, and dramatic singularities at the transitions between regimes. have you looked at how these different regimes and transitions affect your simulations? very different things are happening in each regime, and the cumulative effects are going to give you a unique state at each transition. how would your post-1985 outputs look if you simply fixed historical fire and harvest to the first observed level? or even repeated the observed pattern?

the singularities can cause dramatic shifts in your model outputs due to such large and likely unrealistic changes in disturbance regime. these effects can propagate over time in your simulations and generate large uncertainties in your results.

lines 394-410:

This is confusing. the terms are not consistent and it appears you have redundant terms. at the beginning the 32 tile run is the reference and at the end it is a target. What is a target run (it sounds like one of the 14 simulations you want to evaluate)? what are j model runs? there isn’t any indication that you run the model multiple times for each target simulation.

You don’t have a 32-tile run, unless it is mislabelled in table 3. or it needs to be added to table 3 as the reference run.

equation 11 has n years and i years, and equation 12 has an m index that is not defined.

also, equation 11 isn’t mean square error because your reference is arbitrary (it is a mean square difference); you are not comparing against observations to determine the accuracy of the model.

Results and Discussion

lines 420-421:

this is somewhat due to your input data in that you can have at most one fire event and one harvest event per year.

lines 440-441:

not really. the following two sentences still make sense, though, with respect to the overall response.

line 445:

incomplete sentence

lines 453-466:

what about fire emissions and deforested c (figure 4)? these two variables are quite far off from the increasing tile trajectory.

these are also directly related to your primary goal of simulating fire and harvest disturbance. it would make more sense to optimize for these variable outputs in conjunction with the others, unless you don’t believe that the 32-tile simulation is accurate. tpp should be 6 and rht should be 0.04. especially since the run time is similar across all combinations of these.

or maybe these two shouldn’t be set all because the the 12-tile with these unset has a similar runtime also (but a couple variables have higher difference from 32-tile).

Conclusion

this is a bit redundant with the previous section on implications. these two sections should just be combined for the conclusion.