Dear Reviewer,

Thank you for your thoughtful and constructive feedback on our manuscript. We are pleased that you found the study well-written, with valuable figures and discussions, and we appreciate your recognition of its contribution to the field.

We have carefully considered your suggestion to perform additional analyses, such as using synthetic data to explore observational biases, evaluating model uncertainty at the site level, testing an attention-based model, incorporating lagged features, and adding precipitation as an input feature. While these are all interesting avenues for further research, our study primarily focuses on assessing the efficacy of sequential models for flux upscaling and emphasizing the need for comprehensive experimental setups to identify "best practices" in this domain.

We believe the current analysis addresses our initial research questions effectively and highlights key challenges in ecosystem-atmosphere flux upscaling. Regarding your suggestion about model uncertainty, we agree that the site-level model comparison in Figure 4 would benefit from its inclusion. We have since re-run cross-validation for the top six models from hyperparameter tuning, and will update the figure to reflect model uncertainty (this addition is also presented below).

As for incorporating precipitation as an additional input, we encountered significant challenges. The site-level precipitation data have large gaps, and our down-scaling approach for other meteorological variables is not applicable to precipitation. Using ERA precipitation is also not ideal due to the stochastic nature of precipitation. We believe that the downscaling of precipitation needs further investigation before drawing conclusions, as relying on ERA’s nearest neighbor could lead to misleading results. This method might suggest that precipitation is insignificant when, in fact, the data may not accurately reflect local site conditions. Nonetheless, we would be happy to include this analysis if requested by the Editor.

Below, we provide a more detailed point-by-point response to fully address your suggestions and clarify how they align with or extend the scope of our study.

We sincerely appreciate your understanding and thoughtful approach. Should the manuscript proceed, we will include a comprehensive response to all comments and highlight any changes made in response to your valuable feedback.

Once again, thank you for your insightful input.

Best regards,

Basil Kraft and co-authors

Response to specific comments

L15: Consider adding quantitative information when discussing model performance differences between the sequential and non-sequential models to make the comparative strengths more transparent for readers.

We agree with this suggestion and will include quantitative information on model performance in the abstract to better highlight the comparative strengths of the models.

Lag variables in non-sequential models: Have you considered explicitly adding lagged variables to the non-sequential models? This would offer a more balanced comparison, as it would simulate past dynamics without the complexity of models like LSTM. For example, incorporating lagged climate variables into XGBoost could provide insights into whether sequential models are uniquely beneficial in capturing temporal patterns.

We intentionally did not include lagged variables, as not having to deal with the manual feature crafting is one of the main reasons for using sequential models. Instead, we used proxy variables for system state, such as vegetation indices. While these cannot directly replace lagged meteorological variables, they offer a simpler experimental setup. Designing lagged features for non-sequential models, as highlighted in studies like Tramontana et al. (2016; https://doi.org/10.5194/bg-13-4291-2016), is complex and presents its own challenges.

Model selection – self-attention models: The paper mentions TCN and self-attention as alternatives to LSTM, yet only TCN was tested. Could you elaborate on why the self-attention models were excluded from the experiments?

Based on our experience, attention-based models typically require more training data than what is available in our case, leading us to expect a lower chance of successful training. There are numerous sequential deep learning models (e.g., gated recurrent units (GRU), attention-based models, convolution-based architectures). We chose LSTM and TCN as representative examples. We will clarify this in the revised manuscript.

L127-129: Precipitation appears to be absent from the meteorological variables. Was there a reason for this omission? Precipitation likely impacts ET indirectly through soil moisture, which could be a significant predictor in capturing temporal dynamics.

We agree that precipitation has a significant lagged effect on ET, and its omission is a point that needs further discussion. We will provide additional details in the revised manuscript. While we believe that adding precipitation may improve model performance, particularly for temporal models, downscaling precipitation to the site level remains a significant challenge. ERA observations are suboptimal due to the stochastic nature of precipitation, and this could affect model performance. Given the uncertainties, we are cautious about misinterpreting results. However, we are open to running an additional model configuration with precipitation, should the Editor request it.

L129: For clarity, it would help readers if you briefly explained the significance of the time derivative of potential shortwave irradiation and what processes it represents within the context of ET.

We agree that this information would be helpful. The time derivative is used to help non-temporal models distinguish between morning and afternoon conditions. We will clarify this aspect in the revised version of the manuscript.

L175: The statement "The remote sensing and PFT covariates were repeated in time to obtain uniform inputs" could be clarified. Does this mean daily remote sensing data were kept constant at a sub-daily scale? If so, it’s worth discussing if the model accounts for the sub-daily variations, as metrics like LST and NDWI are not entirely invariant within a day.

Yes, remote sensing variables were kept constant at a sub-daily scale because these variables do not provide diurnal observations. Diurnal variations are driven primarily by hourly-varying meteorological variables, though they interact with satellite-based features that change only on a daily to weekly basis. We will add this clarification in the manuscript.

L204-205: Were the same hyperparameters used for each fold in cross-validation? This clarification would help assess whether the variation within model ensembles arises from differences in training data subsets or distinct hyperparameter settings.

Yes, hyperparameters were tuned once for each model and setup, and they remained constant throughout cross-validation. Thus, the variability within the model ensemble is attributed to differences in the training data and model optimization, rather than variations in hyperparameter settings. We will clarify this point in the revised manuscript.

Fig 4 – PFT impact on TCN and LSTM models: In Figure 4, adding PFT as a predictor seems to penalize TCN performance, yet it enhances LSTM’s accuracy in capturing interannual variability. Could you discuss this divergence?

Considering the scale of the IAV differences (LSTM improves around 0.05 NSE and TCN decreases around 0.01 NSE), this could just be noise from model initialization. Similarly, other setups and scales show small variability which we cannot explain, as for example, the TCN having the same performance in MET and MET+RS on raw time scale, while the LSTM shows an improvement. We argue here that these minor differences should not be interpreted, and we will expand on this in the manuscript. We updated figure 4, as you suggested in the next comment, and indeed the small differences are due to noise.

Fig 4 – Model sensitivity to hyperparameters: Are the displayed results limited to the best models, or could the performance of the other 19 models (with variation bars) be included to show the sensitivity to hyperparameter choices?

The figure 4 displays the test set of the best-performing hyperparameter set per model.  We acknowledge that the aspect of model uncertainty needs to be addressed and we appreciate your suggestion. We ran the cross-validation for the six best models and updated the figure. The updated figure (see below) shows the minimum to maximum performance across the six model runs, and the line represents the best model based on the MSE loss. Note that we do not use all models from hyperparameter tuning, as some models did not converge. From this updated figure, we learn several things, which we will include in the result and discussion sections: 1) As expected, small differences in interannual variability (links to previous question) are due to noise. 2) The xgboost model is extremely robust. 3) The sequential neural networks robustly outperform the non-sequential models on the anomaly scale.

Fig 4—mean-site results: Presenting model performance metrics related to spatial variability, such as the mean site performance, could be informative.

We have added spatial scale to Figure 4 (see the response to the previous comment) and will include a discussion of these new results in the manuscript.

L254-255: While PFT doesn’t enhance site-level predictions, could it mitigate extrapolation errors during upscaling? Including this consideration in the discussion of the scaling-up section may add valuable insight. Or would the spread for the sequential model change with or without PFT?

Our results at the site level suggest that PFTs provide little to no additional value. While there may be minimal improvements on some scales, the potential negative effects, such as unforeseen impacts on upscaling due to the large leverage of PFTs and limited training sites, prevail. We advocate for exploring continuous features (such as plant traits or soil properties) in future studies. This point is discussed in lines 354-363 and 439-444 in the manuscript.

L286-288: Can you test the hypothesis about observation biases with synthetic data or a process-based model simulating extreme events and disturbances? This might strengthen the argument about model vulnerability to changes in predictor distributions. 

We agree with the reviewer that testing this hypothesis with synthetic data or a process-based model could be valuable. However, this is not a primary focus of the study and would require substantial experimentation. Therefore, we will revise our "hypothesis" and present it as a potential explanation rather than a tested claim.

L410-411: Jung et al. (2020) introduced an extrapolation index that might be useful here. Plotting the model spread against this index could demonstrate that model uncertainty correlates with areas requiring more extrapolation, supporting your discussion points.

Unfortunately, applying the method introduced by Jung et al. (2020) is computationally too demanding for our setting (hourly, 0.05degree) due to the need to identify nearest neighbors in the training data for each predicted data point. Adapting this method for large datasets would require significant development and would constitute a separate study. Therefore, we are unable to implement it within the scope of this work.

Dear Reviewer,

Thank you for your thoughtful review and constructive feedback! It’s great to hear that you found our manuscript interesting and see its potential value to the scientific community. Your specific comments and suggestions are very helpful, and we appreciate your time and effort in providing them.

Below, we address your questions and concerns. If the manuscript is accepted and proceeds to the review phase, we will provide more detailed responses, including specific changes made to address your comments.

Thanks again for your valuable feedback,

Basil Kraft and co-authors

Response to specific comments

Line 127–129: Could the authors further clarify how meteorological variable combinations are defined? For example, I noticed that precipitation, which often strongly influences ET, is not included as an input feature in their experiments. Could the authors elaborate this?

We agree that precipitation has a significant effect on ET, and its omission is a point that needs further discussion. We will provide additional details in the revised manuscript. While we believe that adding precipitation may improve model performance, particularly for temporal models, downscaling precipitation to the site level remains a significant challenge. ERA observations are suboptimal due to the stochastic nature of precipitation, and this could affect model performance. Given the uncertainties, we are cautious about misinterpreting results. However, we are open to running an additional model configuration with precipitation, should the Editor request it.

Line 129: How did you calculate the time-derivative of potential shortwave irradiation? Please provide more details.

The time-derivative is calculated as the difference between potential shortwave irradiation values for two consecutive hours, based on the time and location. We will include this information in the revised manuscript.

Line 136: You mentioned each remote sensing product was interpolated to a daily resolution. Did you use nearest-neighbor interpolation, or another method? Please clarify.

This statement is misleading, and we will rephrase it for clarity. Although the MCD43A4 product for the reflectances uses observations from a period of 16 days to characterize and invert the bidirectional reflectance distribution function of a given pixel for the day at the center of the period, this operation is done over a temporally moving window at daily timesteps, resulting in output data with daily frequency. Therefore, no interpolation was required.

Line 184: The phrase “clustering of coordinates” needs further clarification. It would be helpful to include a table (list of sitenames) summarizing the final results of the 8-fold site clustering in the supplementals.

The clustering was performed randomly, but sites within a 0.05-degree threshold were grouped into the same cluster. We will provide a more detailed description of this approach in the revised manuscript and include a table summarizing the 8-fold site clustering results.

Line 191–193: Do you mean that observations from only two random years for each site were selected and used for training, validation, and testing?

We used the full available sequence of observations for training. However, for each training iteration (i.e., each minibatch), a consecutive two-year period was randomly selected from the available data. We will clarify this in the revised manuscript.

Line 294–295: If there are discrepancies between EC site observations and the reanalysis dataset, would it be more effective to build a functional relationship between the reanalysis dataset (e.g., extracted ERA5 values at the site level) and ET observations directly, rather than using site-level functional relationships for upscaling? similar to the methodology in Nathaniel (2023). The authors could consider adding a discussion on this point.

Eddy-covariance measurements of land-atmosphere fluxes are unique and rich in information. The processes involved are highly dependent on local conditions, such as land cover (which can be partially captured by remote sensing features) and weather. By relying on ERA5 features, we would lose access to some of this richness, introducing potential biases and missing high-frequency information from the observations. However, using the same data source for both model training and upscaling could reduce biases, which may offer some advantage. While this is an interesting point, it is beyond the scope of this study. As you suggest, we will elaborate on our choice to use local meteorological observations for model training in the revised manuscript.

Line 376–378: The sentence, “In summary, the sequential models did, when trained on the same subsets of sites, not behave similar in terms of global annual ET. Therefore, it seems unreasonable to assume that the lower global ET estimated by the sequential neural networks is due to a better (and hence more consistent) representation of the processes,” appears ambiguous and redundant. Please revise for clarity and conciseness.

Thank you for your feedback. We will revise this sentence for clarity and conciseness. Our intention is to convey that the temporal models exhibit a small systematic bias, likely because they are more sensitive to shifts in the data source. We will ensure this is more clearly stated in the revised version.