Dear Authors,

Thank you for this interesting study on developing a Machine Learning (ML) based Light-Use Efficiency (LUE) model to estimate Gross Primary Productivity (GPP). The study integrated remote sensing and eddy flux data into a Fast Lightweight  Automated Machine Learning (FLAML) framework and achieved improved statistical performance scores in relation to previous methods. The choice of a ML framework is novel and represents an important step in the development of LUE models.

The study’s methods are sound and statistical analysis very well developed, such tests and choice of training and validation data. However, I have identified some points which could lead to a minor or major revision depending on the editor’s opinion. First, the choice of simple vegetation indices as dependent variables for the model seem to me dated, especially due to the current availability of Solar Induced Fluorescence (SIF) products, which are more suited as proxies of photosynthesis than EVI, NDVI, etc. Although the authors mention the possible future use of SIF, I would like to know further details to why it was not used in this study, or extra analysis where SIF is included. Second, the resolution of the remote sensing products used (500 meters) does not seem to be compatible with the eddy flux data. At this scale, microclimatic or topographic factors may cause significant divergences in relation to a 500 m size pixel, and lead to inconsistencies. I suggest that if possible data with higher resolution are used (LANDSAT or SENTINEL-2) or arguments are given for the use of the lower resolution product. Finally, I would be very interested in the production of a GPP map of China using the FLAML framework, and how it compares with other GPP maps. I think this would greatly increase the manuscript’s appeal.

Specific comments:

L90 - I would not say ML is "fundamentally different" from regression models, but that they offer advantages in relation to.

L94 - I would also point out limitations on ML techniques, such as dependence on large training datasets and not being able to link results to real-world processes

L96 - ...Which is an advantage when the focus is solely on spatial predictions

Fig. 1 - The mini-map on the bottom right corner does not include any sites, or any extra information, maybe remove it? Otherwise, I believe the editors should label these areas in the South China Sea as “under dispute”, as stated in the “maps and aerials” section of the submission guidelines.

Table 2 – In contrast to other vegetation indexes, LAI satellite data is based on empirical models, such as previous GPP estimating methods. It would be interesting to check if field LAI data from the sites are available to see if direct LAI measurements improve the ML model.

L686 - I would argue then that in the future hyperspectral data + ML would provide much better estimates too, this could be discussed with references.

The authors apply FLAML v2.3.3---an automated machine‑learning toolkit---to predict gross primary productivity (GPP) across 20 eddy‑covariance sites, which is a less interesting and less novel endeavor. The manuscript would benefit from a more sharply defined research question and a deeper interrogation of the ecological processes underlying the model's performance. In particular, the authors should clarify what novel scientific insight they seek --- beyond demonstrating sensitivities --- and explore how specific feature groups/selections inform mechanistic understanding rather than merely reflecting data redundancy and uncertainty. Given these substantive concerns about framing and ecological interpretation, I respectfully decline to continue with further review, if that is the case.

- The a whole work reads more like a sensitivity report than a ecological modeling study. What specific scientific insight are the authors seeking by comparing FLAML to not scientifically different feature groups?

- The main text suggest a "FLAML‑LUE model", yet none of the analyses explicitly implement light‑use‑efficiency (LUE) theory. Instead, all results derive from various tree‑based regressors. If the intent is to compare FLAML‑derived machine‑learning against LUE theory, the authors should at least incorporate an explicit LUE model.

– The model groups differ mainly in dryness index definition, data source or temporal averaging (e.g., PDSI vs. evaporative fraction, flux‐tower vs. ERA5-Land temperature, actually Ta_flux is typically gapfilled by ERA5). These inputs often carry overlapping information, so comparisons may reflect data uncertainty or scale mismatches rather than mechanistic differences. Exploring a truly critical predictor --- such as soil moisture --- could strengthen the ecological relevance and offer interesting insights. A basic clarification to mention here is that ERA5-Land is a reanalysis dataset rather than a remote sensing product, and it should not be confused with ERA5. ERA5Land provides hourly rather than daily data.

- The rationale for analyzing 8-day, 16-day vs. monthly statistics is not fully developed. Because GPP seasonality dominates many signals, the differences in model performance may simply reflect sample size (it is unsurprising that monthly R2 exceed those at the 8‑day scale, and this comparison offers no insight)

- Presenting each PFT group in separate sections can make cross‐comparison cumbersome. I suggest grouping figures by PFT (forest, grassland, cropland) with sub-panels for each site or model variant.

- The manuscript contains kind of repeated descriptions across all sessions. I recommend restructuring the whole manuscript thoroughly to avoid duplication. Meanwhile, the authors claim that all results are from validation, but without describing the split strategies. 

This is the review report for “FLAML version 2.3.3 model-based assessment of gross primary productivity at forest, grassland, and cropland ecosystem sites”. The authors developed a FLAML modeling framework to predict vegetation gross primary productivity using hydro-meteorological variables and variables related to vegetation types and elevation. They focus on sites in China and provide detailed model performance in reproducing forest, grass, and crop sites. While the manuscript is structured, the authors need to check throughout the manuscript to ensure readability, e.g., the abbreviations. More importantly, I have some concerns about the modeling data input, cross-validation, and the selection of hydro-meteorological variables. Also, the evaluation of the absolute values of GPP can smooth out potentially poor performance during extreme situations. Therefore, a specific test for stress conditions and an evaluation of GPP anomalies are both highly recommended.

Methodology:

1.GPP and RECO Partitioning:

The manuscript should provide a clear description of the method used to partition GPP and RECO from NEE. Additionally, it is recommended to test different partitioning algorithms to assess their impact on the results.

2.Train-Test Split Strategy:

The procedure for splitting the dataset into training and validation sets needs to be described in greater detail. It is important to test whether the model maintains robustness during stress periods (e.g., droughts or heatwaves). Moreover, model performance should be evaluated not only in terms of seasonal GPP dynamics but also in reproducing GPP anomalies, which are crucial for capturing ecosystem responses beyond typical seasonal cycles.

3.Choice of Environmental Drivers:

The exclusion of key hydrometeorological drivers such as precipitation, vapor pressure deficit (VPD), and soil moisture raise concerns. While LSWI and PDSI are included, they are indirect proxies and not physically direct controls of vegetation water uptake and stomatal regulation. The authors should justify this choice or consider incorporating more directly linked variables.

Specific points:

Line 101: The abbreviation "RFR" should be defined upon its first use for clarity.

Line 134: To promote transparency and reproducibility, the authors should provide a persistent identifier (e.g., DOI) for the datasets used, rather than referencing a general data repository that hosts multiple sources.

Line 178: ERA5-Land should not be categorized as remote sensing data. It is a reanalysis product based on assimilation of observations into a numerical model.

Line 195: The acronym “LSWI” should be spelled out in full the first time it appears.

Table 2: All abbreviations should be clearly defined either in the table caption or as a footnote to enhance readability.

Figure 3 (III): In model evaluation scatter plots, it is more intuitive to place observations on the x-axis and simulations on the y-axis, as this mirrors standard regression analysis practice.

Figures 5/9/13: Do the reported biases account for seasonal differences in GPP variability (i.e., high variability in summer vs. low variability in winter)? Clarifying this would improve interpretation of model performance across seasons.

Figure 7: There is a noticeable underestimation of GPP in DLG (typical grasslands) and overestimation in DXG (alpine meadows). Can the authors explain potential causes for these systematic biases?

Figure 14: Are the farm ecosystems considered in the analysis purely rainfed, or do they include irrigated systems? This distinction is important for interpreting model results under water-limited conditions.