Max Gaber and colleagues investigate the effect of several technical choices in the process of predicting GPP from eddy-covariance measurements and satellite(-derived) data sets using machine learning. The focus is on novel methods in the field of automated machine learning applied to predict monthly GPP at site level from different sets of predictor variables, as well as on the effect of their spatial resolution. The authors demonstrate the applicability of AutoML, and of AutoSklearn in particular, and show that in the global upscaled product, spatiotemporal patterns reasonably compare to other products. They also illustrate the importance of adequate spatial resolution of the predictor variables by increased model performance at site level, when part of the predictor variables are fed into the machine learning at 500m instead of at 0.05deg resolution.

Given the growing number of research studies that implement such data-driven approaches at global and regional scales (large part of whom are cited in the paper) and the still unquantified importance of several technical choices in the set-up, this study is timely and definitely of relevance. The fact that the overall R2 at site level is similar to or slightly higher than from a plane random forest or the results in comparable upscaling exercises at monthly scale (Jung et al. 2011) is interesting, and highlights that tuning the machine-learning set-up may not be the most promising way forward to improving the performance of data-driven models, but rather more informative predictor variables (at least Fig.5 may be interpreted in this way) . I find this a very valuable finding which may also deserve to be communicated/ highlighted more clearly (like you did for example in l. 384-389, but not in the abstract or elsewhere). At the moment the differences in significance are stressed more than the very similar magnitude of performance between the different AutoML methods. Also, your finding that the AutoML does not help to reproduce interannual changes (l. 267) is an important finding, because it is a common problem in data-driven upscaling and very relevant question in the carbon cycle community, and therefore in my opinion deserves to be stressed more.

I suggest publication of the paper after addressing the following major questions/ comments:

1. What is the reason for doing this analysis at a monthly temporal scale when structural vegetation changes dominate rather than finer temporal resolution? I would expect higher gains from AutoML and also more differentiated contributions between predictor variables (especially meteorological features) at higher temporal resolution. This is also the time scale which is more relevant to be able to properly represent seasonal and anomalous trajectories. I would expect large potential from automated model tuning especially for short extreme events, which are relevant for the carbon uptake and hard to represent in a data-driven model set-up, but clearly smeared out at a monthly time step. Much of the discussion in section 4.2 does neglect the coarse time step when for example LUE changes are not expected to play major role.

2. Data sets:

• A number of predictor variables are model outputs themselves, relying on input data and model assumptions. This is not discussed at all.

• What is the reason for ingesting both SIF and instantaneous SIF, or both PAR and RSDN?

• How is the temporal aggregation done?

• How do you handle data gaps?

• Handling of bad data quality is only mentioned for the site-level fluxes, what about the explanatory variables?

• Specify more clearly the data sources, e.g. for the CCI soil moisture, which version did you use? Presumably, FluxCom v6 refers to the FluxCom set up with RSonly ( only satellite-based predictors using MODIS collection 6), which is 8-daily and at high spatial resolution?

3. Spatial resolution: Why not also ingest tower meteorology instead of the coarser ERA5-Land? The scale mismatch could be further       discussed, especially between a 0.05deg pixel and the tower footprint. The way the authors approach the analysis suggests using the 0.05deg pixel is the generally accepted default, which is not the case.

4. In parts the manuscript uses very technical language and describes key concepts only in a very short manner. I suggest to rephrase certain passages to make the manuscript better accessible to a wider audience which may also not be very familiar with the newest developments in the machine learning world – or at least expand more in the supporting information. Examples of very technical sentences in my opinion are l.160-161, l.165-166, l.170-172, l.177-181, l.243-246

5. I am afraid, but I cannot follow the meaning of Fig.6.

Minor comments for clarification:

• Throughout the manuscript: The analysis is not done on climatological time scales, so VPD, precipitation and temperature are meteorological variables, it’s not climate data.

• l.22: I suggest to stress in the abstract already the small differences between the AutoML frameworks, eg. by writing ‘...AutoSklearn consistently but marginally outperformed other AutoML frameworks…’

• l.49 and later in the manuscript: In the literature the term ‘variable importance’ is used with very different meanings. Please clearly state that for your work, importance refers to the contribution of a variable to model accuracy.

• l.49-56: I am not convinced that the conclusions of the different cited papers are strictly comparable because the analyses have been done at different temporal scales, from daily to monthly, and using different feature sets. Although the machine learning results are analysed which do not necessarily need to obey conceptual understanding, the contributions of different features are expected to differ between time scales.

• l.66 (and later as well, eg l.146, 149, 319, 325): Could you clarify/ give examples of what is meant by ‘pipeline creation’ and ‘data processing steps’? The legend of Fig.A2 is hardly understandable for the non-expert without any further context or info.

• l.81: ‘predictive contribution’ to what? To prediction accuracy?

• l. 202: Is there a reason for leaving out the VIs?

• l. 232: So you compute a linear trend also for time series of just 2 years?

• l. 241: What value does the critical difference take?

• Section 3.4: So the main take-away is that the patterns from AutoML in general make sense when compared to other upscaling products? Or do you want to convey another message?

• l.465: the deforestation is mentioned the first time here and I cannot follow what is meant.

• l.519-525: This last part may be slightly overstating, I do not see very clear indications of more robust and accurate GPP predictions yet.

Gaber et al. test the ability of multiple automated machine learning (AutoML) approaches, each based on multiple individual machine learning methods, to upscale gross primary production (GPP) with remote sensing. They specifically test three different AutoML methods (as well as a random forest model as a baseline) with different subsets of remote sensing and meteorological data, finding that they provide very similar performance, with r^2 ranging from ~0.7-0.75 at monthly scale. They also find similar abilities to capture trends, spatial variation, and seasonality across most approaches but that none of them is particularly effective at capture monthly GPP anomalies. The best models were typically based on a combination of MODIS surface reflectance with additional remote sensing-based estimates of LAI/FPAR, land surface temperature, soil moisture, evapotranspiration, and solar-induced fluorescence (SIF); adding meteorological reanalyses of precipitation, temperature, and vapor pressure deficit did not notably improve model performance.

Overall, the manuscript presents an interesting comparison of some cutting edge approaches to automated machine learning and adds a new dimension to ongoing discussions of flux upscaling. It’s also a well written and well-constructed study. The fact that the approaches achieve similar results to each other and to other upscaled products is itself interesting and perhaps suggests that further improvement in upscaled GPP estimates may come from avenues aside from just algorithmic optimization (e.g., better and more extensive ground data, improved remotely sensed data streams). I have a few suggestions for improved presentation and additional analysis, but overall I think this is likely to be a high quality contribution.

General/major comments:

1) My main suggestion for the analysis would be provide, if possible, a more refined and specific assessment of the importance of individual variables. The analysis of the different subsets is interesting, but I think the impact of the study could be enhanced by assessing specifically which variables within those subsets are giving the most “bang for the buck.” I know random forests, for example, provide variable importance metrics and perhaps those are doable from the AutoML approaches as well? I’m curious, for example, in the RS subsets, which variables added the most predictive skill beyond what was achieved with RSmin? How important were LST and soil moisture? Did the ET and SIF data, which are themselves modeled from remote sensing data, add any additional independent information? The CSIF product, for example, is itself an upscaled SIF product based on machine learning of MODIS NBAR data, so it seems like it wouldn’t necessarily add anything beyond what the methods were able to get directly from the NBAR data.

2) I think the Discussion could use a little improvement in places. I think it would be especially helpful to improve how the findings are contextualized in light of previous literature. I’ll provide more specific suggestions below.

3) I find Fig. 6 very difficult to interpret. Is it possible to present those results in a more intuitive form?

Specific comments:

-L12: should that be “scale” instead of “scales”?

-L14: parameterization is misspelled (missing an “e”)

-Fig. 2: Just to clarify, this is showing number of sites, not site-years, correct? If so, I wonder if it would be more relevant to show site-years since that’s a better representation of how much training data is available in each biome?

-L122: I think it would be worth expanding more on these different sources, including references. Especially since some of these (ET and SIF) are themselves modeled based on remote sensing. Given that, what would you expect them to add beyond what would be coming from the NBAR data itself? Would they actually be providing independent information?

-L274-275: These may be “statistically different,” but to me, it seems like an r^2 of say 0.74 is not particularly different from an r^2 of 0.75 in any meaningful sense. The authors do a good job stating this later in the paper, but I do think it’s worth not overinterpreting small differences even if they are “statistically significant.” Any difference, however small, could be “significant” given a large enough sample size, but that doesn’t necessarily make it a meaningful difference.

-L286-297 (but also in other places throughout the results): There are places here that could use references to specific figures or panels within figures. Sometimes it’s hard to tell where the results as described are shown in the figures.

-L304-305: The overestimation of low values and underestimation of high values is interesting and consistent (I think) with some of the early studies of MODIS GPP (perhaps from David Turner and/or Faith Ann Heinsch, if I’m remembering correctly?). Some reference to those earlier works here would provide valuable context. The fact that we’re still trying to solve long-standing problems is itself interesting!

-L390-399: This paragraph (about differences among approaches) seems to slightly contradict the previous one (about how there aren’t really major differences). I’m not suggesting that the authors do a complete rewrite of the paragraph or anything, but I do think it might be worth making sure that they are sending a consistent message: that the differences are generally pretty slight.

-L401-407: It could also be that the quality of the eddy covariance data itself is a limiting factor. EC GPP is used as the ground truth in this case, but it’s not a perfect representation of GPP: EC data has sources of noise and EC GPP is a modeled quantity from the more directly measured NEE. I imagine there may therefore be upper limits to the performance metrics that we can expect when upscaling EC GPP just because of uncertainties in what we’re using as “truth.”

-Section 4.2: I think this section would definitely benefit from a more thorough dive into the variable importance, as suggested in general comments. Also, I don’t think there’s any mention of SIF in this section while other variables composing the RS subset are discussed?

-L433-439: The authors mention this at the end of the paragraph, but I think it could be more up front: reanalysis data (especially for precip) can be very flawed. So maybe temperature and VPD do matter (precipitation probably less so since soil moisture is already included in the model and ultimately it’s soil moisture, not precipitation, that gets directly used by plants) but the reanalysis data just doesn’t do a good job capturing it. Could also be worth a citation to previous literature that has assessed reanalysis data.

-L444: I’d suggest rephrasing “It is to be explored.” That’s somewhat awkward, passive phrasing.

-L463-466: This paragraph is kind of light on citations and the final sentence feels out of place and incomplete, like there’s something more that should be coming that connects the first part of the paragraph to this final thought.

-L477-484: This paragraph is also pretty light on citations. A couple suggestions: Smith et al. 2019 (Remote Sensing of Environment) on challenges specifically in dry regions and the early MODIS papers by Turner that assessed biome differences in MODIS GPP performance. It’d be interested to see the results here contextualized with the challenges that have faced remote sensing of productivity for a long time!

- L481: It’s unclear what’s meant by “high proportion of biomass” or how that would affect productivity estimation. To me, it seems like it’s not high biomass that would lead to good performance but rather high seasonal variation in leaf area (which both DBF and MF have).

-L484: A little unclear what’s meant by “complex biophysical and environmental characteristics.” I think it’d be worth expanding on this and being more specific.

-L487: I think “It is to further research to…” is also somewhat awkward and passive phrasing and would suggest rewording.

-L490: This is another good place to cite Smith et al. 2019, which also shows that drylands are underrepresented in flux networks relative to their global proportion. Haughton et al. 2018 (Biogeosciences) could be a good one too since they showed that drylands are more “unique” (meaning less easy to apply a globally-trained model to an unseen site) than most other systems, which may be partly why the underrepresentation of dryland sites in flux networks can be such a problem for upscaling in those regions.

-L504: For the Conclusions section, it might be worth expanding on what’s meant by “RS” here. That’s referring to a specific subset of the variables but for readers who are skimming and skip to the conclusions section, they might miss what that subset refers to.

-L519-520: Maybe to some extent, but it’s interesting to note that RF (not automated and with, I think, some amount of subjectivity in choices) performed nearly as well as the AutoML methods.