Improving numerical snowpack simulations by assimilating land surface temperature

Alonso-Gonzalez et al.

Review by Bertrand Cluzet

General Statement

In the context of upcoming satellite missions imaging the land surface temperature (LST) with unprecedented spatial and temporal resolutions, Esteban Alonso-González and his colleagues investigate the potential of their novel, versatile data assimilation toolbox, MuSA, to assimilate such observations in order to improve energy-mass balance simulations of the snowpack. This is an exciting and promising avenue, and their effort to investigate snow surface temperature, arguably a blind spot of the snowpack modelling community, should be acknowledged, and strongly encouraged. I must add that this short paper was very pleasant to read and fits well within the scope of the Cryosphere, and want to apologize to the authors and editorial board for the delay in delivering this review.

Snow surface temperature is indeed a key variable of the snowpack surface energy balance. For instance, surface temperature observations have been long used as a boundary condition to close SNOWPACK model’s  energy budget under freezing conditions [1]. When the snowpack surface reaches the melting temperature, the surface energy budget can no longer be closed, but observations can be used to detect (or reject) the occurrence of surface melt, which is a valuable information per se. Finally, LST values above freezing point inform us about the absence of snow on the ground.

This study is an OSSE (Observing System Simulation Experiments), where the authors simulate future LST observations from the upcoming TRISHNA mission with ‘degraded’ model outputs from FSM. They then assimilate such observations in ensemble simulations also using FSM (therefore, this is an ‘identical twin’ OSSE), accounting for precipitation and air temperature uncertainties only. Their 4 study sites span a wide diversity of snowpack conditions from alpine to arctic climates. The authors investigate the influence of cloud coverage on the availability of observations by assimilating these observations under different fictive cloud coverage scenario. Their results suggest that Land Surface temperature has the potential to constrain snow water equivalent with astonishingly good performance, in particular when it comes to Landsat-like low revisit time (see e.g. Fig. 2 and l. 215-216). This seems ‘too good to be true’ as the authors provide little evidence to nuance their results (l. 239-243). Landsat data has been out for decades now, how could the snow data assimilation community really miss something so promising for that long, or is this result somewhat overestimated? The authors discuss this point, but the reader may remain sceptical and I see several reasons why the performance would be rather overestimated.

Identical twin OSSEs are very tricky to set up and should be designed with a lot of care for reliable conclusions to be drawn [2]. Unfortunately, the setup in this study does not convince me for a reason that requires to delve into some developments. Because (1) they do not integrate any snow model parameter perturbations, (2) do not confront their simulation with real (in-situ or satellite) observations of LST (and potentially SWE), and (3) do not discuss the literature on snow surface temperature modelling by snowpack models such as FSM, their OSSE implicitly lays on the fundamental hypothesis that FSM can accurately model both the snowpack’s surface temperature and the SWE. This ’perfect model’ assumption is a typical shortcoming of  identical twin OSSE’s, that usually results in overly optimistic conclusions on assimilation performance (and other pitfalls) as thoroughly discussed in Sec VI.2.2 of [2].

In fact, some elements in the current literature show that models similar to FSM [3], [4] (or more sophisticated [5], [6]) can exhibit rather strong (negative) surface temperature biases with respect to in-situ measurements, despite representing bulk snow properties such as the snow water equivalent (SWE) with a satisfying accuracy. As a user of FSM, I know that very accurate SWE simulations can be obtained with more than 2-3K of negative surface temperature bias (I am pleased to provide the authors with evidence for that).

As an example, a physical process that could lead to an overestimated impact of surface temperature assimilation on SWE is the liquid water routing through the snowpack. Qualitatively, the snow melts near the surface, and is routed towards the bottom, implying percolation, retention, and refreezing (in particular), before it runs off (SWE decrease). FSM represents water routing with a bucket approach, which is a very simplified representation of the truth (see e.g. [7]). Because only one version of FSM is used in the current setup, there is a common ‘transfer function’ from surface melt (surface temperature) to runoff (SWE ablation), between the synthetic truth and the ensemble members from the assimilation run. A good surface melt (surface temperature) gets artificially mapped into the good runoff (SWE decrease). But the truth is for sure blurrier, and there are plenty of plausible runoff scenarii associated with a given surface melt scenario: in real life, assimilating surface temperature might result in looser constraints on the SWE.

Therefore, in the current setup which does not allow the data assimilation algorithm to update FSM parameters (and therefore adjust liquid water routing, or tackle error compensations with albedo, surface density, turbulent fluxes within FSM), data assimilation of real (and accurate) observations, will probably perform much worse than the presented results. As an example, assuming a negative surface temperature bias from FSM, correcting surface temperatures (by increasing them) by way of data assimilation will very likely degrade SWE modelling performance. This is arguably a bigger potential problem, than the question of cloud influence on observation availability, and should therefore be assessed (not necessarily addressed) first.

I am confident that the authors can address the above comments, but nevertheless, I think that they are substantial enough to require a major rework of the study: the conclusions of this study rely on strong hypotheses that are not substantiated enough, and significantly reduce their applicable range. Furthermore, these hypotheses could be partially avoided, but at the cost of a substantial redesign of the study. I am please to suggest pathways to address these major comments. I would further suggest that the authors add a detailed discussion on remaining shortcomings of their approach as listed below and in the annotated manuscript.

A premise for this OSSE experiment is therefore to (a) compare FSM outputs with in-situ or satellite surface temperature observations. I am confident that this can easily be achieved, thanks to the amazing versatility of MuSA, that would allow them to run it at the FluxAlp station [9], or any of the sites within the ESM snowMIP setup (I think that data is publicly available) [10], for example. In the absence of biases, or any potential ‘mismatch’, (b) in-situ or real Landsat observations [9] could even be assimilated as a proof of concept, as a very nice addition to the current study. (c) In the presence of biases, the current study should be substantially redesigned, probably by integrating snow model parameter perturbations, in order to allow the snow model to ‘learn’ from its surface temperature errors, and trying to fit more into the identical twin OSSE guidelines provided in [2].

(d) An additional easy (thanks to FSM modularity) and minimal (although still quite imperfect) requirement for this OSSE would be to use a different FSM configuration for the generation of the synthetic truth (as acknowledged l.239-243) to reduce the inbreeding of the current identical twin setup.

Below, and throughout the annotated manuscript, the authors will find other minor (but substantial) comments that would also need to be addressed in a revised version of the manuscript.

Minor comments

1. As the authors say, LST informs on the snow surface temperature as well as on snow absence (when LST>273.15K). In the smoothing mode, because information can be propagated backwards in time, it is hard to tell apart which one of these pieces of information have the most impact, although the poor performance of the filter points towards a significant impact of the melt out date info on the smoother performance (as discussed in l. 209-214). Companion DA experiments discarding LST observations past the melt out date (or stopped in the core of the winter season) could help tell this apart more rigorously. This would be essential for real-time applications where there is no information available about the future melt-out date.
2. This paper provides no evidence of the actual behaviour of LST and SWE timeseries (there are for SWE, but only the mean, and for different cloud coverage scenarios, not for a single run). Adding example plots with openloop/synthetic truth/ assimilation runs including ensemble spread would be insightful.
3. No information is given on the physical configuration of FSM used in this study. In particular, it would be important to know whether the albedo is parameterized as a function of surface temperature.
4. An undiscussed question is the potential (tiny) mismatch between what is modelled (the skin temperature of the snowpack, assuming a flat infinite surface) and what the satellite will see : the thermal infrared emission of a potentially rugged snow surface within a mixed pixel. I expect the difference to be tiny, but adding a discussion sentence on this would be appreciated.
5. The Snow surface temperature assimilation being rather scarce, please consider citing [11] which also assimilated SST although in a multivariate setting.

Technical notes

• The title does not reflect the content of the paper. ‘OSSE’, or ‘synthetic’, or ‘towards the assimilation of…’ or ‘feasibility’ should reflect the fact that no real data is being assimilated.

References:

[1]        M. Lehning, P. Bartelt, B. Brown, T. Russi, U. Stöckli, and M. Zimmerli, “SNOWPACK model calculations for avalanche warning based upon a new network of weather and snow stations,” Cold Reg. Sci. Technol., vol. 30, no. 1, pp. 145–157, 1999, doi: 10.1016/S0165-232X(99)00022-1.

[2]        B. K. W. Lahoz and R. Menard, Data assimilation. Berlin, Heidelberg: Springer, 2010. doi: https://doi.org/10.1007/978-3-540-74703-1.

[3]        C. B. Menard et al., “Scientific and human errors in a snow model intercomparison,” Bull. Am. Meteorol. Soc., vol. 102, no. 1, pp. E61–E79, 2021, doi: 10.1175/BAMS-D-19-0329.1.

[4]        G. Arduini et al., “Impact of a Multi‐Layer Snow Scheme on Near‐Surface Weather Forecasts,” J. Adv. Model. Earth Syst., vol. 11, no. 12, pp. 4687–4710, Dec. 2019, doi: 10.1029/2019MS001725.

[5]        I. Gouttevin et al., “To the Origin of a Wintertime Screen-Level Temperature Bias at High Altitude in a Kilometric NWP Model,” J. Hydrometeorol., vol. 24, no. 1, pp. 53–71, 2023, doi: 10.1175/JHM-D-21-0200.1.

[6]        M. Lafaysse, B. Cluzet, M. Dumont, Y. Lejeune, V. Vionnet, and S. Morin, “A multiphysical ensemble system of numerical snow modelling,” Cryosphere, vol. 11, no. 3, pp. 1173–1198, 2017, doi: 10.5194/tc-11-1173-2017.

[7]        N. Wever, C. Fierz, C. Mitterer, H. Hirashima, and M. Lehning, “Solving Richards Equation for snow improves snowpack meltwater runoff estimations in detailed multi-layer snowpack model,” The Cryosphere, vol. 8, no. 1, pp. 257–274, 2014, doi: 10.5194/tc-8-257-2014.

[8]        B. Cluzet, M. Lafaysse, E. Cosme, C. Albergel, L.-F. Meunier, and M. Dumont, “CrocO_v1.0: a particle filter to assimilate snowpack observations in a spatialised framework,” Geosci. Model Dev., vol. 14, no. 3, pp. 1595–1614, Mar. 2021, doi: 10.5194/gmd-14-1595-2021.

[9]        A. Robledano, G. Picard, L. Arnaud, F. Larue, and I. Ollivier, “Modelling surface temperature and radiation budget of snow-covered complex terrain,” The Cryosphere, vol. 16, no. 2, pp. 559–579, Feb. 2022, doi: 10.5194/tc-16-559-2022.

[10]      G. Krinner et al., “ESM-SnowMIP: Assessing snow models and quantifying snow-related climate feedbacks,” Geosci. Model Dev., vol. 11, no. 12, pp. 5027–5049, 2018, doi: 10.5194/gmd-11-5027-2018.

[11]      G. Piazzi, G. Thirel, L. Campo, and S. Gabellani, “A particle filter scheme for multivariate data assimilation into a point-scale snowpack model in an Alpine environment,” The Cryosphere, vol. 12, no. 7, pp. 2287–2306, Jul. 2018, doi: 10.5194/tc-12-2287-2018.

I have included all my comments in the attached pdf. There is also an anotated version of the pdf. My anotations are in French (sorry, I did not initally plan to include this) but I believe at least one author speaks French.