Liang et al. describe the development of a global 1km data product for land surface – air temperature differences and sensible heat fluxes. In principle such data-driven products are very valuable for the community. However, I have several methodological concerns, primarily related to the validation approach and with respect to variable selection as input to the models.

Major points:

• The cross-validation strategy chosen by the authors is not adequate and yields overoptimistic results. It is absolutely compulsory to stratify train and test data by sites and not (only) by time. This is simply because data from one site are not independent and the objective of the study is to estimate at unmeasured locations. This needs to be done correctly.
• The authors chose Rn and ET as input to the model to predict H. In my opinion this is hard to justify as H=Rn-LE-G and predicting ET is a similar problem as predicting H. I would find it conceptually more appealing if input variables are close to observations and not already derived products with additional layers of uncertainty.
• The authors chose slope and aspect as predictors. While it is clear that these variables are very relevant in principle, the footprint of flux towers is supposed to be restricted to reasonably flat terrain. Therefore, I cannot imagine that robust patterns wrt these terrain variables can be learned.
• The authors also chose day of year as predictor, which has no direct environmental meaning. I suggest to drop this or replace by e.g. potential radiation or sun angle.
• The authors mentioned a ‘circular’ approach between training and testing for hyper-parameter tuning (line 300). It is absolutely forbidden to use test data for any kind of model tuning. Perhaps this is a misunderstanding. Please clarify.
• The authors use the Twine et al approach to correct flux tower based sensible heat fluxes by forcing energy balance closure. This is a critical assumption, which needs through discussion because the uncertainty related to energy balance correction is very large, esp. for H (see Mauder et al 2024, AFM)

Minor points:

• I find the uncertainty estimates listed in Table 1 and referenced in the text a bit misleading as they are not comparable among the products because they were not calculated consistently
• Model tree ensembles for FLUXCOM mentioned in table 1 is likely wrong as I suppose the authors used the ensemble product
• Line 113: sentence starting with “Therefore” seems incomplete
• The choice of LSTM for estimating H is unclear – have not seen a clear comparison to RF and the other methods (did I miss this?)
• Are the comparisons of global H values in section 5 based on exactly the same spatial domain. This matters as e.g. FLUXCOM does not cover deserts where H is particularly large.

It is a good idea to directly estimate daily land surface-air temperature difference (Tsa) and sensible heat flux (H) using machine learning method.

The manuscript explores the linkage between Tsa and its predictors and demonstrates the feasibility of establishing this linkage using machine learning.

The performance of the developed product is comprehensively evaluated and superior to that of corresponding reanalysis products and satellite product.

I expected to see the complete product soon.

Detailed comments,

1. How to remove the effects of spatial resolution mismatch between multi-source predictors on the estimated Tsa and H?

2. If the estimated Tsa is used to derive the physically based model, what is the accuracy of H.

  This study presents two valuable global datasets spanning 2000-2020: land surface sensible heat flux (H) and land surface-air temperature difference (Tsa), generated using data-driven approaches. Recognizing that Tsa is the primary driver of H, the authors first refine the estimation of Tsa and subsequently employ it to predict H. Overall, the article is well-structured, with detailed explanations of the algorithm design, variable selection, and comparative analyses against existing products across multiple scales. The generated datasets hold significant value for global energy balance studies.

  A notable limitation is the absence of data for 2021–2024. Additionally, a few minor issues should be addressed:

1.  Table 3: the last row, there is a spelling mistake on “Leaf area index”;keep the initial letter case consistent for variable names in the first column of the table.
2.  InSection 2.2.2, the GLASS product provides ET and FVC data at an 8-day temporal resolution, whereas the model requires daily input. Please clarify how this temporal discrepancy was addressed (e.g., through interpolation,or another method). A detailed explanation should be included for reproducibility.

Comments to essd-2025-136

Considering that accurate estimation of Tsa is a fundamental prerequisite for accurately estimating H, the authors first employed a RF to estimate Tsa, followed by the estimation of H based on LSTM, ultimately producing a high-resolution dataset. This work is of significant value for studies on climate change and land–atmosphere interactions. However, the Methods section is overly detailed and somewhat cumbersome, and the Results section lacks clear organization. Therefore, the manuscript still requires further improvement before it can be considered for publication.

I would like to offer the following suggestions, which I hope will useful for the authors.

Introduction

It is recommended to consider changing the notation from "T_sa" to "T_s–a".

The authors emphasize the importance of Ts–Ta in estimating sensible heat flux (H), which is understandable. However, since the H product has already been produced, it remains unclear why there is still a need to derive or produce Ts–Ta separately. The manuscript does not sufficiently justify the necessity of generating Ts–Ta as an independent product, especially given that H has already been estimated. A clearer explanation of the added value or specific applications of the Ts–Ta product is needed to support its relevance.

The manuscript does not provide a clear explain for selecting Random Forest (RF) and Long Short-Term Memory (LSTM) as the baseline methods for predicting H and Ts–Ta. The choice of these specific models requires further justification.

Methods

The methodology section is overly detailed. It is recommended to streamline the description to enhance clarity and readability.

It is recommended to introduce the data used for model development and those for comparison respectively. Presenting these two types of data separately will help improve the clarity of the manuscript.

Result

The superiority of the Tsa dataset is primarily illustrated through comparisons with other available products. While this approach is valuable, it is recommended to include a more thorough and explicit discussion of the advantages of the generated Tsa data.

Estimating Tsa as an intermediate step before deriving H may provide more reasonable than directly estimating H. A comparative analysis of these two approaches would be valuable in highlighting their respective strengths and limitations.

The title of section 4.1 is “Uncertainty quantification of Tsa model”, however, this section mainly compared product Tsa with GLASS and ERA5-land.

P245： Please remove the validation method to Method section.

It is noted that Figure 7 is missing tick marks for the R² axis. Including these would enhance the figure's readability and allow for better quantitative comparison.

Although comparing with other daily H datasets is valuable, the discussion spans eight paragraphs, which may overwhelm the reader. Condensing this section while retaining the key findings would improve readability and better highlight the strengths of the proposed dataset.

Discussion

To enhance readability and guide the reader through the key arguments, the Discussion section would benefit from the inclusion of subheadings that reflect its main themes.

While the analysis of variable importance in estimating Tsa is useful, the primary objective of the study is the accurate estimation of H. Therefore, investigating the contribution of each variable to the estimation of H would be more directly aligned with the study’s goals and would provide greater practical insight.

The current manuscript outlines the application scenarios of Tsa; however, since H is the final product of interest, its application scenarios should also be discussed. This will better demonstrate its scientific and practical significance.

Conclusions

Please consider refining and condensing the description of both the methods and the main conclusions. This will help to emphasize the mainly point of the manuscript.