See attached document.

Review of

A comprehensive land surface vegetation model for multi-stream data assimilation, D&B v1.0

by Knorr et al.

General comments:

This paper sounds like a well written technical report and needs major changes before it can be published in GMD as a model description paper. A number of models, not all very complex, are mature enough to do the same job (and more) as the one proposed by the authors. Why is this new model needed? In terms of process representation, I could not see any innovation in the proposed modeling framework, except for the simulation of SIF and VOD. I assume that the innovation is in the data assimilation part, but no example is shown. Since data assimilation is not demonstrated here, "data assimilation" should be removed from the title. Similarly, data assimilation is mentioned frequently throughout the paper, but no example is provided. Data assimilation could be mentioned briefly in the Introduction and as a perspective in the Discussion section. No more. A clear definition of data assimilation is also lacking. Data assimilation can be done in many different ways. As far as I could understand, in this paper data assimilation is equivalent to "model parameter tuning". This is quite different from the variational or sequential Kalman filtering methods used in meteorology and in some land modeling frameworks to initialize initial conditions (e.g. of root-zone soil moisture) at a given time. This should be clearly explained. It should also be explained why the authors do not trust their default model parameter values. I assume that these default model parameter values come from the scientific literature. Why not trust them? Finally, a critical risk of parameter tuning is that the tuned model might be good for bad reasons, which is unacceptable for a model that aims to explicitly represent the main biophysical processes ("process-based modeling system"). This is acknowledged by the authors on L. 546. But how do the authors ensure that this does not happen? This is not clear. Finally, the rationale for "parameter tuning" is that improved static model parameter values are needed for the surface component of climate models and for climate change impact models. I do not believe that the current model is designed for such applications. What is the real purpose of the model? For monitoring and reanalysis, sequential assimilation would be preferable to model tuning. This work combines the BETHY and DALEC models. A comparison of the model simulations is presented over two contrasting sites (Spain and Finland). Why these two sites in particular? Results from the comparison are not good, which tends to show that this new model is not a good model. Or maybe these sites are particularly difficult to represent? Could you indicate score values from other models over these sites?

Recommendation: major revisions.

Particular comments:

- L. 54-59 (list of requirements): Is it something that other models could not do?

- L. 97 (daily time step): I believe that the daily time step is not sufficient to represent snow processes. Especially when snow melt occurs.

- L. 154 (potential photosynthesis): It is not clear whether potential photosynthesis varies from one day to another according to solar radiation and leaf temperature. Could you clarify?

- L. 275: The ICOS data portal contains a large number of sites. Why have you selected these two sites in particular?

- L. 324 (overestimate of GPP): Over the Boreal site, GPP is much more than "overestimated". There is nearly a factor of two at summertime. Why is the model that bad?

- L. 433: On L. 214, Ssif = 1 and on L. 434 Ssif = 10. Could you explain why?

- L. 481: "in the simulations, soil moisture decreases to near zero": why? Could this be caused by the overestimation of soil evaporation (a classical modeling problem)?

- L. 486-487 (carbon fluxes [...] are simulated reasonably well): You cannot say that for the Boreal site.

- L. 529: The large uncertainty on Ssif shows that the biophysical basis for SIF is very weak in the proposed model. Why not using machine learning to build an observation operator for SIF?