text/comments-and-submissions/2020-08-20-istem.md

In ashiklom/edr-da: ED2 RTM testing functions

Hi Alexey (and others),

I gave the manuscript a closer read, please find my comments below. I have no major issues, but while the main narrative is cutting out the middle-man, I wish we could have had an actual comparison where we show what happens when you do the same work without cutting out the middle man (using MODIS LAI for example). Also, lack of out-of-sample validation is a bit concerning.

But overall, I think there is enough material that warrants reporting as a paper, also I think it satisfies GMD criteria. Massive amount of work, congratulations Alexey. Feel free to decide what to include or not these are all suggestions/questions, and let me know if I can help in any way.

Hope these help, Istem

L13: Could we be a little more specific in the last sentence of the abstract? "Modeling surface reflectance" is a bit vague. E.g. something along these lines "..that having dynamic vegetation models directly predict surface reflectance is a promising avenue for their calibration and validation using remote sensing data."

L22: Intro assumes the reader knows what calibration is, why it is necessary, why it is hard and how remote sensing could help with some of the challenges. Given the journal, I'm fine with this decision, but just wanted to point it out.

L35: I would remove "effectively" to highlight our approach

L50: You may want to come back to this point in discussion.

L59: There is a missing reference (?)

L64: You may want to be consistent between ED and ED2 usage

L69: I would add "jointly" already here to avoid misunderstanding that calibration is done individually. E.g. "we jointly calibrate this model at 54 sites" or "we calibrate this model at 54 sites jointly"

L70: Before the last sentence of the intro, this is a good place to mention your hypotheses. What do you expect from the calibration? Any prior expectations taking the EDR structure / study sites / joint calibration into account?

Y vector: Was the element before the last supposed to have Sn instead of S1?

L95: Eqn 4 doesn't have Sn+1 but Si+1, could the text also mention Si+1 instead?

Eqn 6: There are 2 m2i,2i, was the latter one meant to be m2i+1,2i?

L164 (also later on for priors): It would be nice to give these parameters in a table, e.g. parameter name, definition, unit, prior form.. also grouped by category maybe leaf traits, canopy traits

L173: I think the paper requires a supplementary table (a csv?) thats list the species and PFT assignments explicitly

Figure 1: Are there more than 54 sites in the main figure? Is it because some sites have multiple obs? It would also be nice to mention it here in the caption again. Also didn't we have some fancier versions of this graph? (e.g. attached) But I'm fine with this version also.

Eqn 23: why not annotate posterior also?

L194: Does the subscript r refer to reflectance?

L210-212: Reduced equifinality, could be good to come back to in the discussion.

Figure 3: If not all sites are necessary (i.e. if there is no additional information in reporting all of the sites) maybe highlight 3-4 cases and move this to the Appendix. Even if there is some interesting additional information, that should be distilled into a separate graph, e.g. overall PROSAIL versus EDR performance

L259: What about density? Also were the tests just visual? Have you conducted any other exploratory analyses? Also not sure, but at a glance calibration seems to be doing better at Late Hardwood (A8) and Pine sites (A9) than others, and there happen to be more LH and Pine sites than the others, is it a matter of more information fed into the calibration about these types?

L260: Was there an improvement with respect to the uncalibrated version? - Didn't look at this

L261: I feel like Fig 5 is also showing signs of miscalibration, no? A pattern I would expect from a joint calibration. Not sure if Mike would agree. - Not sure I understand. Disregarding for now.

Discussion reads uncoupled from the results. Maybe discuss the superiority of the approach with respect to "not cutting out the middle man" approach a bit more strongly. You mentioned a couple of studies in the intro (L25-27) that use data products, what were their struggles and which solutions does your approach have to offer? Could bringing models to data rather than the other way around remedy these issues for example: https://bg.copernicus.org/articles/17/3733/2020/ - Partially addressed...

I would also discuss the results of doing the joint calibration (doing equally good/bad everywhere) and maybe mention the hierarchical way. - Skipping for now...

PRO4SAIL is not mentioned in the discussion at all, what have we learned from that comparison? - Dropped SAIL from the analysis

ashiklom/edr-da documentation built on April 16, 2021, 9:33 p.m.