General comments

The paper by T. Malle et al. addresses the effects of spatial resolution, atmospheric forcing data, and land surface characterisation on snow depth, gross primary production (GPP) and evapotranspiration (ET) simulated by the Community Land Model version 5 (CLM5) over Switzerland. Factorial model experiments with different combinations of resolution and input data were used to isolate the effects of the three sources of variability in the model output. The authors conclude that deficiencies in all three aspects contribute to the current uncertainty in the results of land surface models (LSMs), particularly in heterogeneous regions like the Swiss alps. In this context, they call for the use of more fine-grained input data and for evaluating LSM simulations at high resolution.

The study contains several elements that are interesting for the audience of ESD, including modelers and users of land/climate model output over heterogeneous regions. It probably involved huge efforts, considering the generation of various input datasets, model experiments, the analysis of three simulated quantities (snow depth, GPP, ET), and their comparison against observational or modelled benchmarks. For a high-quality modelling paper, I think the technical foundations need to be improved (e.g. what controls simulated snow depth, GPP, and ET in CLM5-SP offline?). Also, the study addresses several complex issues in land/climate modelling (e.g. ecophysiology, carbon cycle dynamics, feedbacks, etc.) that are poorly captured in the model setup chosen for this study (i.e. CLM5 with prescribed vegetation phenology, largely inactive carbon and nitrogen cycles, and prescribed atmospheric conditions). I think this distracts from the relevant key findings and confuses the overall storyline. Below, I list suggestions how the study and its findings could be improved to become more useful for the audience of ESD. I think this involves substantial revisions. My comments refer to main text only, not the Appendix.

Specific comments

(1) Relevance of higher resolution and implications of the results. It is obvious that higher resolution and more specific input data improve model performance for a specific location, i.e. when CLM5 output is evaluated against high-resolution simulations or point/site-scale observations. This is known by the modelling community and it is thus not surprising that the 1km simulations with high native-resolution input data perform best. What additional insights can the study provide, e.g. should we focus on (1) high resolution atmospheric forcing data in offline land surface modelling, (2) high resolution land surface data, or (3) high LSM resolution to capture non-linear effects between atmospheric forcing and land surface characterisation? Or should model evaluation and benchmarking be done at high resolution, to rule out deficiencies due to insufficient resolution and focus on process uncertainties? What do users get if they run CLM5 at high (<0.5°) resolution with CRU as the atmospheric forcing dataset, is CRU 0.5° simply interpolated and could “smart downscaling” considering the temperature lapse rate be useful as a (built-in) solution? (this would be very interesting for the CLM community!)

(2) Benchmarking and implications. The purpose of a land surface model like CLM5 is to simulate larger spatial scales ranging from one grid cell to regional domains or the globe, in spatially representative grid cells rather than point fashion (only at very high resolution, grid cells start to resemble point/site-scale conditions…). Therefore, it might be useful to evaluate the output of various simulations (1km, 0.25°, 0.5°) at the lowest resolution of 0.5°, to assess if there are major differences due to input data quality and non-linearities. Such results could inform the modelling community if there is a need to account for small-scale heterogeneity to obtain accurate fluxes and pools at larger spatial scales of one to several kilometres. This might ultimately be more relevant for the purpose of CLM5 (or LSMs in general) than mimicking point/site-scale conditions.

(3) Land surface dataset methodology and evaluation. The generation of a high-resolution land surface dataset based on national land cover data and satellite-based LAI (not sure I interpreted the brief description in section 2.3.2 correctly) seems quite innovative and interesting. I think the procedures should be described in more detail, so that other could potentially follow a similar approach. Some summary/evaluation/validation beyond what is shown in Figure 1 (e.g. how did land cover fractions and LAI change across Switzerland, regionally, or for the point locations?).

(4) Overall scope including snow depth and GPP/ET. The manuscript might benefit from limiting the scope and analysis to snow depth, and in that area developing the causes for improved model performance more thoroughly (e.g. how can spatial resolution, atmospheric forcing data, and land surface characterisation influence snow depth considering forced precipitation, rain/snow partitioning inside CLM5, land cover, LAI, slope, etc.). A caveat of this is, of course, that snow depth is likely closely linked to the forced precipitation and temperature fields, which could make the results appear trivial. Yet, the authors might identify interesting aspects related to, e.g., land cover, LAI or slope. In any case, there are multiple reasons why the analysis of effects on GPP and ET should be excluded, or included only after substantial revisions:

a) The link between Hypothesis 2 and the chosen methods is currently very weak. Most importantly, the modelling setup and the correlation analysis do not allow to isolate the effects of differences in snow depth from differences in its drivers (i.e. spatial resolution, atmospheric forcing data, and land surface characterisation) on GPP and ET. The framing of GPP and ET as “snow-cover dependent ecophysiological variables” is confusing, because likely most of the GPP and ET differences are driven by the resolution-dependent forcing fields directly (i.e. atmospheric variables or land surface characterisation) and not indirectly via snow-cover changes. This joint independent driver could lead exactly to the correlation between snow cover and GPP/ET found by the authors, without any effect of snow cover itself on GPP/ET.

b) In CLM5 in SP mode, vegetation phenology is prescribed as a climatological seasonal cycle of LAI, and LAI controls the leaf to canopy scaling of all fluxes including carbon (GPP) and water (ET). Therefore, in SP mode the model has very limited capabilities to show an ecophysiological response to snow cover change, in the sense of seasonally shifting growth. As a minimum, I recommend to discuss the precise implications (e.g. is the response we see solely due to changes in temperature and water availability, or what can affect GPP in SP mode at all?). However, I strongly doubt that an ecophysiological response can be quantified in SP mode. I imagine it works as follows, although I am not 100% sure: if the snow season is shorter than in the climatology, suitable temperatures for growth coincide with zero LAI in the model and nothing happens until the climatological growth begins; if the snow season is longer than in the climatology, suitable temperatures for growth coincide with high LAI in the model, leading to a jump start and no compensation later in the season.

c) Ecophysiology refers specifically to (plant) organisms and not to land as a whole. So, for ecophysiological effects on ET I suggest analysing plant transpiration and canopy evaporation fields of CLM5, or total ET in the vegetated parts of grid cells excluding the bare soil PFT. However, considering that the experiments include changes in land surface characterisation, the best option might be to remove the term “ecophysiological” and to refer to responses in land ET, which include e.g. the effects of varying bare soil fraction.

d) Simulated GPP and ET are compared against the “best-effort” configuration of CLM5 at 1km resolution as a benchmark, and deviations from this benchmark are considered model uncertainty. An objective benchmark (e.g. observational data or output of a dedicated model, like for snow depth) is lacking. Also, I am not sure the term “model uncertainty” is appropriate in this context. Is it model uncertainty if lower resolution models perform poorer at producing high-resolution-model-like outputs? The comparison could potentially be done the other direction (see point 2).

(5) Introduction: The background could be more technical (e.g. what controls simulated snow depth, GPP, and ET in CLM5-SP offline?) and focused on aspects that matter for this study (i.e. strengthen the main story and cut out things that are interesting but not directly relevant).

(6) Discussion: The findings need to be contextualised, considering the capabilities of CLM5 in SP mode offline (see above). For the comparison with other studies (e.g. Birch et al.) to be valid and useful for the reader, it would be good to understand what these studies did or how they explain their biases (e.g. also land cover specification, atmospheric forcing and model resolution, or something completely different like effects of grazing animals or artic plant types?). Certain links made between the findings of this study (with a very specific setup and research focus) and model uncertainties (with sometimes known different root causes) are not appropriate (e.g. L371). I recommend reconsidering these and focusing the discussion more on new insights gained through this study (see points 1 and 2).

(7) Language and figures: The manuscript is very well written. There are a couple of “empty” phrases that highlight something but do not actually deliver new insights (e.g. the results have profound implications, the study highlights the importance of model development, the study highlights the utility of multi-resolution modelling, etc.). I think those could be filled with content or removed. The figures have high quality and are visually appealing.

Technical and line-by-line comments

L14 ff: “Earth’s systems” sounds a bit unconventional and far-reaching; maybe use a more concrete/narrow term?

L19: add water

L23, L24: maybe use “influence” instead of “control/determine”

L24, L35, L65: there is no feedback among the mentioned dependencies/effects

L29, L31: check logical link for “thus” and “as”

L49: offer literature for multi-resolution modelling?

L50: they allow evaluating

L51: I guess any gridded LSM can be evaluated against point/site-scale observations by taking the individual grid cells that match the locations best; there is no need for dedicated point simulations for this

L56: what is meant by snow cover “dynamics”? the temporal evolution?

L73: consider limitations in SP mode

L80: in my understanding, “process representation” refers literally to how processes are represented in the model, i.e. the equations used to calculate snow depth, GPP and ET; I think this term is not appropriate for the modifications in forcing/input data and resolution made in this study

L87: remove “heat fluxes” if not addressed in the results

L88-90: consider reformulating, the sentence sounds nice but it does not deliver any content (or at least I cannot understand it)

L98-102: the methods need some technical precision here: which state variables, which datasets (e.g. what time period/ past conditions do they represent), not only “natural” vegetation (crop PFTs), how does GPP work in SP mode (if the LUNA model active, also cite Ali et al. 2016 for photosynthesis)

L100: “is approximated”: I think this a used choice, not done by the model

L103: revise components of ET, e.g. soil sublimation does not make sense; maybe “ice” is missing?

L105-111: snow cover is the focus of this study, so I think the foundations of snow cover calculations should be provided for convenience (and understanding), including rain/snow partitioning in CLM5

L115, L197: PFTs are patch level, mention that prescribed LAI etc. is in SP mode

L116-L119: is there any difference between taking out individual grid cells from the regional simulations and running dedicated point simulations (i.e. a 1x1 regional grid) in your setup? they should be identical, provided that (1) the grid is anchored and specified identically, and (2) there are no lateral exchanges between grid cells (which depends on the CLM5 compset, if river routing is off there is no lateral exchange)

L124: for comparing between resolutions, it would have made sense to subdivide the 0.5° and 0.25° grids, e.g. by using 4x the number of cells for 2x resolution, i.e. 20x12 grid cells for 0.25°; that way you could preserve the grid anchoring and preclude differences due to a new “positioning” of the grid; maybe motivate your choice and/or mention potential effects of different grids on the results

L128: a “the” is missing

L137 ff: are the grid cells for point simulations centered on the station coordinates? using the “nearest neighbor” grid cell for land surface characterization seems like a very simplified approach compared to all the other sophisticated things done in this study; why not use conservative regridding so you would get something more realistic? ideally, one would generate surface data from the raw PFT fraction data at 0.05° resolution (CLM’s own methods) or the raw Swiss national data (your methods); in contrast, taking the nearest neighbor effectively shifts surface information and pairs it with the correctly positioned atmosphere; if this is acceptable, one might as well take the nearest (or interpolated/regridded) results of the gridded simulation? (see comment on L116-L119)

L154: is “accelerated decomposition” valid/applicable for SP mode? it sounds like BGC; by “cycling” (remove “re-“)

Figure 1 caption: is “percentage vegetation cover” the natural vegetation landunit including bare soil, or the sum of vegetation PFTs and CFTs? (the latter would be good)

L159, L177: is CRU a station-based interpolated dataset and the OSHD based on the COSMO model? for OSHD, it is also a bit unclear if the dataset was produced or re-used for this study

L187: was the native 0.05° PFT and LAI data reprojected and regridded, or was this done based on an existing surface dataset at e.g. 0.5° resolution? depending on the data used, there might be several regridding steps involved (with every step further degenerating the final product) and the native input might be 0.05 or 0.25°; was this done with the CLM5 tools, with which regridding algorithm (bilinear, conservative)?

L200: FSM2 output is not “observational”

L216-217: remove “were”, “ground truth” is usually used in remote sensing?, does “upscaled” mean regridded and if so, with which algorithm (bilinear, conservative)?

L230-234: consider doing this at 0.5° resolution, see point 2

L238: to be honest I am a bit lost by now – was 3.1.1 done with the point simulation results? maybe for each Results sub-section this could be highlighted in the title or mentioned in the first sentence

L247: for this section it would be really good to understand how land cover and LAI and potentially affect snow depth in CLM5 (see point 4)

L283: replace “parameters” by “variables”

L284: why is peak GPP assessed and not total GPP? I see the “motivation” later in L328, but because there is a bigger effect does not mean it is more relevant? I think this is related to limitations in SP mode (see point 4b); the effects described in section 3.2 go way beyond ecophysiology (see point 4c)

Figure 4 labels: replace “climatological” by “meteorological”

L322: snow cover and ET are negatively correlated, but I doubt this is driven by snow but rather by cold temperatures and energy (not water) limitation (see point 4a); feedbacks to the atmosphere are missing in CLM5 offline by construction

L348 ff: for the calculation of variations in (monthly) total GPP across Switzerland, it would be useful to have an observational benchmark and to relate the amounts to total GPP (i.e. % variation of total GPP). Is there a good reason for not calculating variation in total annual GPP? I would find this quantity more informative

L368: ET can also be water limited in Switzerland, at least in some regions seasonally

Although land surface modeling has evolved from simple biophysical parameterizations to complex frameworks in recent years, large uncertainties remain especially in mountainous regions and areas with complex terrain. This study uses a multi-resolution modeling setup to investigate the impact of meteorological forcing data, spatial resolution, and land surface data on the simulation of snow cover and ecophysiological variables. The authors perform simulations using the Community Land Model version 5 (CLM5) over Switzerland. They found that increased resolution not only improved the representation of snow cover in CLM5 but also propagated through the model and affecting the gross primary productivity (GPP) and evapotranspiration (ET).

Overall the manuscript is well written and of interest to the land surface/earth system modeling community. However, the CLM5 model setup and the model evaluation method do not appear to be appropriate in the current manuscript. In specific, (i) the CLM5 is setup in prescribed satellite phenology mode (~ fixed growing season), yet one of the main focus of the study is to investigate the link between snow cover duration and growing season length (and GPP/ET); (ii) CLM5 simulations in the study were conducted at three resolutions: 1km, 0.25deg, and 0.5deg, but the evaluations were performed at 1km resolution (Figure 3), which are not fair comparisons in my opinion.

I suggest the authors choose the prognostic biogeochemistry mode for CLM5 simulations, and perform model evaluation at the resolutions of the respective simulations. In addition, I have some minor suggestions outlined in the comments below that will hopefully improve the future version of the paper.

Specific comments

L164, the sentence does not read well.

L169-170, do you assume values in the original dataset (ClimCRU1km) are at sea level, and apply the temperature lapse rate based on the mean elevation of each 1km grid from the CRUJRA data? Which elevation data do you use? I would suggest include a map of elevations in Figure 1.

L172, it would be helpful to add a description of the snow/rain partitioning method in CLM5.

L181-182, do you just aggregate the 1km data to 0.25deg and 0.5deg? Please try to describe what exactly is being done. The ClimOSHD forcing data would be useful for other modelers, is the data available?

L234, given that the met forcing and landcover data etc. are all at coarser resolutions, it is not fair to evaluate coarser resolution (0.25deg, 0.5deg) CLM5 simulations using finer resolution (1km) observations. I suggest the authors regridding the 1km observation data to the 0.25deg and 0.5deg first, then redo the comparisons and Figure 3.

L241-245, it would be helpful to show or discuss which variables in the met forcing data contribute to the different CLM5 simulations.

Figure 3 is an important figure in the paper, but the Taylor plots and labels/legend are too small, and hard to read.

In the captions of all the figures, a summary of main results is also included, which is not necessary and makes the captions too long.

 L256, supplementary material is not found.

L265, Figure 3 needs to be cited here.

L273-275, I suggest the authors redo these evaluations at the resolutions used for each CLM5 simulations.

L290-291, the sentence doesn’t read well.

Figure 4, note the 3^rd panel are labeled as effect of climatological forcing instead of meteorological forcing.

L316-317, the sentence does not read well.

The manuscript by Malle et al. investigates the impact of spatial resolution, quality of atmospheric forcing datasets and land-use information on the simulated snow depth, GPP and ET over the spatial extent of Switzerland and adjacent watersheds of neighboring countries by using the Community Land Model 5 (CLM5). Simulations of different combinations of meteorological forcing and land-use information were conducted to analyze changes in model performance. In addition, CLM5 simulated snow depth were compared with station observations and results from a spatially distributed, physics-based snow model. The authors find the combination of increased spatial resolution of model and high-quality input datasets can improve the representation of snow cover in CLM5, and these improvements further propagate through the model, directly affecting GPP and ET. The manuscript demonstrates the importance of high spatial resolution and high quality input datasets for climate impact studies.

The manuscript dedicated a detailed description of methodology, but the explanation of the results is somewhat brief, and most of them are descriptive, lacking of model processes related analysis and discussion. Such as, what controls the snow depth simulation in CLM5, how the different forcing datasets affect snow simulation? how the improvements in snow propagate in CLM5 in a cascade way, what’s the linkage between snow cover and GPP and ET. I suggest the authors improve these parts. In addition, the figures in the manuscript should be improved. e.g. Figure 3 & 5 are too small and hard to read.