Review Reports - AI-Driven Weather Data Superresolution via Data Fusion for Precision Agriculture

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript addresses a relevant and critical issue in the field of precision agriculture and applied climate services: the generation of spatially detailed and operational meteorological information at the field scale, capable of capturing microclimatic variability, based on operational coarse-resolution numerical weather prediction products. The study proposes an innovative approach based on the fusion of global climate data (GFS), observations from agroclimatic stations, and static physiographic descriptors, integrated using machine learning models and a hybrid spatial super-resolution scheme supported by KNN, with a clear focus on its operational deployment in sensor infrastructures.

The document is well structured and presents a coherent methodology, accompanied by an evaluation using spatial-temporal divisions that allow for analysis of the models' generalizability. The results demonstrate consistent improvements over the reference model, providing a viable, practical, and scalable technical solution to bridge the gap between low-resolution operational meteorological products and the high-resolution needs of modern agriculture. However, in order to increase the clarity, methodological transparency, and scientific rigor of the work, as well as to deepen the interpretation of the results and address their inherent limitations more explicitly, I consider it pertinent to address a series of major observations and comments, as well as some of a methodological nature, which will strengthen its scientific contribution, reproducibility, and comparability with other studies before considering its publication.

General comments

C1. What are the advantages and disadvantages of this study?

C2. Although the objectives (H4) state that physiographic descriptors are “necessary” to capture location-dependent biases, no experiments or results are presented on this.

C3. Please, look at your Discussion, is there a real comparison to other researchers of your results?. It is necessary to carry out a thorough comparison, I recommend to include some relevant references, in order to improve the discussion on the novelty of your study, comparing to the others.

*The answer to these questions should be reflected in the manuscript*

Specific comments

Line 67: Change “in situ” to “in situ.” Do the same throughout the manuscript.

Lines 97-110: The literature review on meteorological ML is adequate, but the authors should add at least one recent reference (2023–2024) on ML-based post-processing of NWP to reinforce its relevance.

Lines 87-110: The transition to advances in machine learning is appropriate, but the list of methods (LSTM, U-Net, CatBoost, etc.) could benefit from a clearer structure that explicitly links each methodological category (deep neural networks, gradient boosting, spatiotemporal architectures) to the specific problem of meteorological or agricultural super-resolution. Currently, it seems a bit like a list.

Lines 132-135: According to whom? Cite relevant literature.

Line 182: The term “significantly” is mentioned without any formal statistical proof being presented later. It is recommended to refine the language or add statistical analysis.

Line 187: The authors should define more explicitly what is meant by “spatial pattern reconstruction.”

Line 212: Figure 1 needs improvement. It is important for the authors to add the regional and global context of where the region of interest is located. It is also important to add some notable places, contour lines, or a thematic layer showing the topography of the area.

Lines 223-228: Explicitly indicate the exact lead time of the GFS used (00 UTC +24 h, 06 UTC +24 h, etc.), as this aids reproducibility.

Line 239-246: A more detailed description of the specific criteria for excluding “erroneous” data or data with “sensor problems” should be added.

Lines 242-246: What about the stations in terms of altitudinal representativeness? The authors should include a table showing these characteristics (Lines 267-272).

Line 244: What is the number of stations for training and validation?

Line 274: SHAP is mentioned in previous work, but no SHAP results are shown in this manuscript. Correct.

Line 309: The figure is not mentioned in the text.

Line 309: The figure does not reflect the “Overall System Architecture.” The authors should generate a more detailed general flowchart.

Line 320-325: The mathematical formulation requires clearer notation. In addition, clarification is needed on whether the models predict anomalies or absolute values.

Line 342-354: The authors should discuss the limitations of TabPFN's scalability and dataset size.

Line 352: Same as Line 244.

Line 361-364: The decision to set K as the number of available stations deserves a more solid justification. Were other values of K explored? A brief sensitivity analysis or a reference to an optimization criterion (e.g., cross-validation in the feature space) would strengthen this choice.

Lines 361-364: The authors should add a conceptual comparison with kriging or IDW, even if they are not implemented.

Line 394: The authors should add at least: RMSE (sensitivity to extremes), Bias (mean error), Daytime/nighttime analysis.

Line 401-Table 1: No confidence interval or significance test is reported. Include a nonparametric test (Wilcoxon or bootstrap) between GFS and TabPFN–KNN.

Line 434: Since there is no continuous spatial “ground truth,” qualitative assessment is necessary. Figures 5 and 6 are crucial to this section. The manuscript indicates that they are used to assess topographic gradients and spatial patterns. However, the authors should explicitly mention in the text which specific spatial features (e.g., valley-summit temperature gradient, slope aspect effect) the models are expected to capture and whether they succeed, based on visual inspection. In addition, the authors should calculate thermal gradients vs. altitude and analyze excessive smoothing (variance loss).

Author Response

Dear Reviewer 1,
We sincerely thank you for your thorough review and constructive comments. We appreciate your positive assessment of the relevance, operational focus, and methodological coherence of our work. Following your suggestions, we revised the manuscript to improve clarity, methodological transparency, reproducibility, and the interpretation of results. Below, we provide a detailed point-by-point response.

General comments

C1. What are the advantages and disadvantages of this study?

Author response: We have explicitly added a “Strengths and limitations” paragraph at the end of Section 4 (Discussion), immediately before Section 5 (Conclusions) (highlighted in yellow in the revised manuscript; approx. lines 526–535). This added text summarises the main advantages (operational end-to-end workflow combining GFS predictors, station observations, and physiographic descriptors; consistent improvement over raw GFS, with TabPFN–KNN achieving MAE 1.26 °C vs 1.66 °C) and the key disadvantages/limitations (station-based validation without spatially continuous ground truth; restriction to one region and a 24 h horizon; potential smoothing of localised extremes by the KNN propagation).

C2. Although the objectives (H4) state that physiographic descriptors are “necessary” to capture location-dependent biases, no experiments or results are presented on this.

Author response: We agree that the original wording of H4 was not sufficiently supported by explicit results. We therefore (i) revised H4 to avoid an unconditional “necessary” claim and to clarify how the role of physiography is evaluated (highlighted in the Research goals and hypotheses section; approx. lines 192–194), and (ii) added a quantitative analysis in Section 4.2 (Interpretation of the KNN superresolution effect) together with Table 2 (highlighted; approx. lines 507–514). Table 2 reports the MAE reduction obtained by the physiography-conditioned KNN propagation (ΔMAE = MAE_base − MAE_base–KNN), showing that the largest improvements occur when stations are withheld (space=validation), consistent with physiography-informed similarity helping to mitigate location-dependent biases

Author response: We agree that the original Discussion did not sufficiently benchmark our findings against related work. We therefore expanded the Discussion by adding a dedicated subsection “4.5. Comparison with related work and novelty”, where we (i) compare the achieved error reduction relative to raw NWP to recent ML-based NWP post-processing studies, and (ii) position the contribution with respect to recent AI/precision-agriculture and crop-yield literature, clarifying the novelty of producing an operational, high-resolution temperature product as meteorological forcing for downstream agronomic decision-support. Relevant new references were added and cited (Refs. [43]–[48], appended at the end of the reference list). The inserted text and new references are highlighted in the revised manuscript.

Specific comments

Line 67: Change “in situ” to “in situ.” Do the same throughout the manuscript.

Author response: Formatted in situ as italic.

Author response: We incorporated the requested recent (2023–2024) references on ML-based post-processing of NWP into the literature review around Lines 97–110 by adding a short follow-up passage immediately after the sentence ending with ref. [25] (now Lines ~111–119 on page 3). The added sentences are highlighted in yellow. We also appended the corresponding new references as [39]–[42] at the end of the reference list

Author response: We revised the literature review paragraph in Lines 87–110 to provide a clearer structure. The methods are now explicitly grouped into (i) deep neural networks for spatial super-resolution/downscaling, (ii) gradient-boosted tree ensembles for station-level post-processing and bias correction, and (iii) spatiotemporal architectures for temporally consistent post-processing, with each category directly linked to the super-resolution/post-processing problem addressed in this study. The revised text starts with “From a methodological perspective, recent ML work…” and is highlighted in the revised manuscript.

Lines 132-135: According to whom? Cite relevant literature.

Author response: We supported the statement that many existing systems are limited to station-level forecasts or coarse-grid NWP products by adding explicit citations at the end of the sentence (highlighted) and adding the corresponding references to the reference list as Refs. [49–51].

Line 182: The term “significantly” is mentioned without any formal statistical proof being presented later. It is recommended to refine the language or add a statistical analysis.

Author response: We revised the wording of the hypothesis to avoid implying formal statistical significance. In H1, we replaced “significantly” with “consistently”, and in H2, we replaced “statistically significant” with “consistent and practically meaningful”. Additionally, in Section 2.4.4, we clarified that the study reports descriptive error metrics across samples and split regimes and therefore discusses effect sizes without formal significance testing. The revised text is highlighted.

Line 187: The authors should define more explicitly what is meant by “spatial pattern reconstruction.”

Author response: Spatial pattern reconstruction is the process of inferring a spatially continuous, high-resolution field from sparse point predictions or coarse inputs by propagating learned values across space using geographic or physiographic similarity. The text was changed accordingly.

Line 212: Figure 1 needs improvement. It is important for the authors to add the regional and global context of where the region of interest is located. It is also important to inlcude notable places, contour lines, or a thematic layer showing the topography of the area.

Author response: Added elevation and cities as local context and topography, and mentioned the area name for the global context.

Lines 223-228: Explicitly indicate the exact lead time of the GFS used (00 UTC +24 h, 06 UTC +24 h, etc.), as this aids reproducibility.

Author response: Added to text: We select only 24-hour lead time predictions (00 UTC + 24h, 06 UTC + 24h, 12 UTC + 24h, 18 UTC + 24h) for simplicity, but other lead times can be used just as well.

Line 239-246: A more detailed description of the specific criteria for excluding “erroneous” data or data with “sensor problems” should be added.

Author response: The description is added in the text: implausible temperature extremes or abrupt discontinuities inconsistent with local meteorological conditions

Lines 242-246: What about the stations in terms of altitudinal representativeness? The authors should include a table showing these characteristics (Lines 267-272).

Author response: Added Figure 2 to show the distribution of altitudinal and slope representativeness.

Line 244: What is the number of stations for training and validation?

Author response: Added to text: We selected 30 stations for training and 6 stations for validation.

Line 274: SHAP is mentioned in previous work, but no SHAP results are shown in this manuscript. Correct.

Author response: We agree. The previous wording could be interpreted as if SHAP results were produced in this manuscript. We revised the sentence to clarify that the feature selection was guided by the SHAP/feature-importance analysis reported in our earlier work [15], but that we do not repeat the SHAP analysis and therefore do not present SHAP plots in the current manuscript. The revised text is highlighted in the updated version.

Line 309: The figure is not mentioned in the text.

Author response: Added mention of figure to 2.3.1

Line 309: The figure does not reflect the “Overall System Architecture.” The authors should generate a more detailed general flowchart.

Author response: The architecture scheme was added as Figure 3

Line 320-325: The mathematical formulation requires clearer notation. In addition, clarification is needed on whether the models predict anomalies or absolute values.

Author response: Edited the text to: The super-resolution task is formulated as a supervised regression problem. For each station, a feature vector is constructed from GFS predictors, recent station observations, and static physiographic attributes. Using these inputs, the model predicts the two-meter air temperature at a lead time of twenty-four hours. The objective is to minimize the mean absolute error (MAE) between the predicted and observed station temperatures. The models are trained to predict absolute temperature values rather than temperature anomalies.

Line 342-354: The authors should discuss the limitations of TabPFN's scalability and dataset size.

Author response: Added to text: However, TabPFN does not scale efficiently to large datasets. We therefore trained it on a random 10% subsample of the data, which may limit its ability to exploit rare regimes. Its performance should thus be interpreted as indicative rather than fully representative of full-data training.

Line 352: Same as Line 244.

Author response: Added to text: We selected 30 stations for training and 6 stations for validation.

Author response: Added to text: The optimal number of neighbours is set to the number of available stations, which empirically yielded the best results, minimizing MAE on the validation set.

Lines 361-364: The authors should add a conceptual comparison with kriging or IDW, even if they are not implemented.

Author response: The information was added in the text:
Conceptually, the proposed physiography-conditioned KNN propagation is related to classical spatial interpolation methods such as inverse distance weighting (IDW) and kriging. Unlike IDW, which relies solely on geometric distance, the present approach operates in a multidimensional feature space that includes elevation and terrain derivatives. In contrast to kriging, which assumes a parametric spatial covariance structure, the KNN layer is nonparametric and data-driven, making fewer stationarity assumptions while remaining computationally lightweight for operational deployment.

Line 394: The authors should add at least: RMSE (sensitivity to extremes), Bias (mean error), Daytime/nighttime analysis.

Author response: Added RMSE table

Line 401-Table 1: No confidence interval or significance test is reported. Include a nonparametric test (Wilcoxon or bootstrap) between GFS and TabPFN–KNN.

Author response: Included Wilcoxon and bootstrap test results between GFS and TabPFN–KNN.

Line 434: Since there is no continuous spatial “ground truth,” qualitative assessment is necessary. Figures 5 and 6 are crucial to this section. The manuscript indicates that they are used to assess topographic gradients and spatial patterns. However, the authors should explicitly mention in the text which specific spatial features (e.g., valley-summit temperature gradient, slope-aspect effect) the models are expected to capture, and whether they succeed, based on visual inspection. In addition, the authors should calculate thermal gradients vs. altitude and analyze excessive smoothing (variance loss).

Author response:

We have rewritten section 3.4 and added section 3.5:

3.4 Spatial superresolution maps: qualitative assessment

Since spatially continuous ground truth for 2 m air temperature at the target resolution is not available, a qualitative assessment based on visual inspection is necessary. Figures 5 and 6 are therefore used to evaluate whether the superresolved fields reproduce physically plausible and topography-driven spatial patterns. In particular, the assessment focuses on (i) monotonic temperature gradients with elevation (valley–summit contrasts), (ii) enhanced cooling in valley bottoms indicative of cold-air pooling, and (iii) asymmetries related to slope orientation and local terrain gradients.

Visual inspection indicates that all learned models reproduce large-scale elevation-dependent temperature structure more consistently than the raw GFS baseline. Hybrid KNN-based configurations produce spatially smoother fields with coherent temperature gradients aligned with terrain elevation. Valley regions generally exhibit lower predicted temperatures than surrounding slopes, consistent with expected nocturnal cold-air drainage, although the strength of these minima varies across models. Slope- and aspect-related temperature contrasts are partially captured, but tend to be attenuated in the KNN-propagated fields, suggesting a trade-off between spatial smoothness and the representation of localized microclimate extremes.

3.5. Quantitative analysis of spatial gradients

To complement the qualitative assessment, we analyze predicted temperature gradients with elevation as a simple quantitative proxy for spatial realism. For each model, temperature–altitude relationships are estimated by regressing predicted grid-cell temperatures against elevation across the domain. The resulting lapse-rate-like gradients are compared between models to assess whether physically plausible cooling with height is preserved.

In addition, spatial variance of predicted temperature fields is evaluated as an indicator of excessive smoothing. Models employing KNN-based spatial propagation exhibit lower variance than standalone backbones, confirming that the interpolation step enforces spatial continuity while also suppressing small-scale variability. This variance reduction is most pronounced in complex terrain, where localized cold-air pooling and sharp thermal contrasts are expected. These results are consistent with the visual patterns observed in Figures 5 and 6 and highlight the superresolution layer's smoothing behavior.

Reviewer 2 Report

Comments and Suggestions for Authors

The article proposes a workflow that applies the basic activities of machine learning systems, such as ingestion, preprocessing, training, and publishing. In particular, the main goal is to predict 24hs ahead 2m air temperature and generate spatially continuous high-resolution temperature fields.

In general terms, the workflow acquires and ingests data sources from external forecast data (GFS) and station measurements (regional station observations from southern Moravia), and preprocesses them by merging dynamic predictors with static physiographic attributes (elevation and terrain gradients). Next, the workflow trains data by applying different model families. These models are evaluated under identical spatiotemporal cross-validation schemes. In particular, the workflows applies LightGBM, TabPFN, Transformer, and Bayesian Neural Fields. Next, the spatial mapping is implemented via a KNN interpolation layer in physiographic feature space.

The article is well-written and structured; however I have some concerns regarding the organization, the previous works and the validation. The first concern relates to previous work. In particular, the authors explain that the study presented here is an extension of reference [15]:

Koutenský, F.; Pihrt, J.; Cepek, M.; Rybář, V.; Šimánek, P.; Kepka, M.; Jedlička, K.; Charvát, K. Combining Local and Global Weather Data to Improve Forecast Accuracy for Agriculture. Unpublished manuscript, based on experiments with HadISD, GFS, and ERA5-Land for the Czech Republic. DOI:10.15439/2024F5990.

The work appears as unpublished in the references, but the DOI indicates that it was published in:

Annals of Computer Science and Information Systems, 2024, Vol.41, p.77-82

is it the same article?

Furthermore, it is not clear what extensions have been made to that work (presented in [15]) nor the different approaches of both proposals. Apparently, in this new article, data sources were changed (from [15]) and now the goal is to predict the air temperature near the surface; whereas in [15] the study combined data from local weather stations with global forecasts (GFS) to improve local weather predictions. So, why do the authors consider this work as an extension of [15]? it seems that both studies pursue different objectives, with different data sources and model predictions. In [15] the authors compared the performance of CatBoost, MLP, and LSTM machine learning techniques against baseline models raw GFS predictions, and the last measured values from the stations). However in this work, the workflow was created to compare the performance of seven models (BayesNF, LightGBM, TabPFN, Transformer, LightGBM–KNN, TabPFN–KNN, Transformer–KNN) against the baseline GFS. So, where is the extension? And how and why do these amchine learning techniques were chosen?

I do not understand Figure 2. What are the subfigures? The figure is not referenced in the text nor explained. What is the ALIANCE infrastructure? Reference?

Other concern relates to the five hypotheses defined. While they may be correct and useful in the context of the article, the authors only defined them at the beginning of the article without providing any further explanation. These hypotheses were not analyzed at the end of the study to determine whether they could be verified or not. The Discussion and Conclusions sections did not consider these hypotheses, so the validation of the workflow remains incomplete.

I consider that the authors should include a more complete explanation of Figures 5 and 6, describing each of the subfigures shown and their interpretations.

Finally, are the data and code used in this proposal available? Can other researchers use them for the presented or new comparative analyses? Is the experiment fully reproducible?

Author Response

Dear Reviewer 2,

We sincerely thank you for your careful reading of our manuscript and for your constructive feedback. We appreciate your positive assessment of the writing quality and the overall structure. Following your suggestions, we revised the manuscript to clarify the relationship to our previous work, improve figure explanations and system description, explicitly address the hypotheses, and strengthen reproducibility and availability statements. Below, we provide a point-by-point response.

The work appears as unpublished in the references, but the DOI indicates that it was published in:

Annals of Computer Science and Information Systems, 2024, Vol.41, p.77-82

is it the same article?

Author response: Corrected, text was published

Furthermore, it is not clear what extensions have been made to that work (presented in [15]), nor the different approaches of both proposals. Apparently, in this new article, data sources were changed (from [15]) and now the goal is to predict the air temperature near the surface; whereas in [15] the study combined data from local weather stations with global forecasts (GFS) to improve local weather predictions. So, why do the authors consider this work as an extension of [15]? it seems that both studies pursue different objectives, with different data sources and model predictions. In [15] the authors compared the performance of CatBoost, MLP, and LSTM machine learning techniques against baseline models raw GFS predictions, and the last measured values from the stations). However in this work, the workflow was created to compare the performance of seven models (BayesNF, LightGBM, TabPFN, Transformer, LightGBM–KNN, TabPFN–KNN, Transformer–KNN) against the baseline GFS. So, where is the extension? And how and why do these amchine learning techniques were chosen?

Author response: We revised the statement describing the relationship to our previous study [15] to make the extension explicit (highlighted in the Introduction paragraph immediately before “Research goals and hypotheses”). We now clarify that, while [15] addressed station-level NWP post-processing, the present manuscript adds an explicit spatial superresolution layer to produce spatially continuous high-resolution fields, and extends the evaluation by (i) introducing physiography-conditioned KNN propagation, (ii) benchmarking additional, more recent model families (LightGBM, TabPFN, Transformer, BayesNF) and their KNN-hybrid variants, and (iii) testing generalization in both space and time. In addition, we added a short model selection rationale in Section 2.4.2 (Base Predictive Model) (highlighted), explaining why these model families were chosen and how they relate to operational constraints and complementary inductive biases.

I do not understand Figure 2. What are the subfigures? The figure is not referenced in the text nor explained. What is the ALIANCE infrastructure? Reference?

Author response: Figure 2 was referenced, and the schema of the ALIANCE architecture was added as the current Figure 3

Author response: We agree. To explicitly validate the hypotheses, we added a dedicated subsection “4.5. Evaluation of hypotheses” in the Discussion, where H1–H5 are revisited and linked to the corresponding quantitative evidence (Tables 1–2) and relevant sections. We also added a concise summary of the hypotheses at the end of Section 5 (Conclusions). All additions are highlighted in the revised manuscript.

I consider that the authors should include a more complete explanation of Figures 5 and 6, describing each of the subfigures shown and their interpretations.

Author response: Added a more detailed description to Figures 5 and 6 (now 7 and 8).

Finally, are the data and code used in this proposal available? Can other researchers use them for the presented or new comparative analyses? Is the experiment fully reproducible?

Author response: Data are available through Agriguv and Focal Hubs, codes are part of commercial IPRs.

Reviewer 3 Report

Comments and Suggestions for Authors

The article is devoted to the development of a method to map of air temperatures at 2 m (24 hours in advance) with high spatial resolution. To do this, the authors combined data from the standard meteorological monitoring of regional stations in South Moravia (Czech Republic) and the global weather forecast GFS (0.25°), as well as static physical geographical descriptors. To create the most effective models, authors used a number machine learning methods (LightGBM, TabPFN, Transformer, and Bayesian neural fields), as well as special interpolations with the KNN spatial interpolation method. As a result, the TabPFN–KNN scheme turned out to be the best with an error margin better than GFS.

Despite the interesting results obtained, there are the following comments :

1) Undoubtedly, the air temperature at 2m is an important parameter for agriculture. However, it is not clear how a short-term 24-hour temperature forecast at high-resolution grid (in isolation from other parameters such as air humidity, wind speed, and precipitation) can be used in practice of precision farming, in particular, for frost protection.

2) Since in the article there are no data of the surface area covered by 36 weather stations in South Moravia (Czech Republic), it is not possible to estimate their density.

3) To understand the local characteristics of the locations of these 36 weather stations, it would be desirable to present (graphically for example) the elevations and aspects of their locations.

4) Similar difficulties arise due to the lack of data of the high-resolution pixels generated for the numerical experiments.

5) To compare the both characteristics (elevation and aspect) of the weather stations and generated pixels, it is desirable to present the corresponding distribution functions, as well as to compare them. This made it possible to divide pixels into two categories – inside and outside boundaries established by characteristics of weather stations as well as use them for result analyses.

6) The proxy functions used include only the characteristics of static physiography. However, the type of land use and land cover during the growing season often significantly affects the air temperature. The reason of this indicator exclusion has not found in the article.

7) It is not clear how, on the basis of which considerations, the spatial resolution of the DEM with sub-kilometer scale (500 m) was set according to ASTER data, which has a significantly better scale, and how the elevation and aspect were calculated for stationary weather stations (lines 248-255).

8) It is not clear why the K-Nearest Neighbor method was chosen for spatial interpolation.

9) There are no results or references of publications on the conducted preliminary analysis with feature importance analysis based on CatBoost and SHAP values (lines 275-278).

10) It is not clear for which time period the results presented in Table 1 and Figure 3 were obtained.

11) It is not clear for which weather stations the results are presented in Figure 4.

12) Figures 5 and 6 show maps of predicted air temperatures for cases A and B. However, there is no description of them. Moreover, it is not clear from these figures how the presented results differ, and help to understanding the influence of microclimatic characteristics.

Author Response

Dear Reviewer 3,

We sincerely thank you for your careful reading of our manuscript and for your constructive comments. We appreciate your positive evaluation of the proposed workflow and the achieved improvements over the GFS baseline. Following your suggestions, we revised the manuscript to clarify the practical relevance, improve the description of the station network and spatial grid, justify key methodological choices, and enhance the interpretation and documentation of the presented results. Below, we provide a point-by-point response.

Author response:We revised the Introduction paragraph beginning “These developments are closely connected to precision agriculture …” to clarify the practical role of short-horizon (hours to one day) high-resolution 2 m temperature forecasts for frost protection. The added text explains that decisions are primarily driven by the expected nocturnal minimum temperature and its timing (i.e., whether/when to activate mitigation and allocate resources), and that high-resolution fields are particularly relevant for resolving within-field gradients due to cold-air pooling (added sentences highlighted; see Refs. [49–51]).

2) Since in the article there are no data of the surface area covered by 36 weather stations in South Moravia (Czech Republic), it is not possible to estimate their density.

Author response: Added approximate area size to Figure 1 description.

3) To understand the local characteristics of the locations of these 36 weather stations, it would be desirable to present (graphically, for example) the elevations and aspects of their locations.

Author response: Added elevation to Figure 1.

4) Similar difficulties arise due to the lack of data of the high-resolution pixels generated for the numerical experiments.

Author response: Updated Figure 1 should also help with this.

Author response: Added Figure 2, which shows all the distributions

Author response: We agree that land use/land cover conditions during the growing season can affect near-surface air temperature. In the current manuscript, the proxy functions and similarity features were intentionally limited to static physiographic descriptors (elevation, slope, aspect, and land-cover class) because the proposed workflow is designed for operational near-real-time inference with low-latency inputs, whereas dynamic vegetation/crop-state indicators (e.g., NDVI/LAI or crop phenology layers) are not consistently available at the required spatial/temporal resolution and latency for our operational setting. To address the comment, we (i) revised RG2 to explicitly state that the evaluated geographic contribution is based on static descriptors only, and (ii) added a clear explanation of the exclusion of time-varying LULC/crop-state features (and stated this as a limitation and future extension) in Section 2.2.3 (Static Geographic Data). The added/modified text is highlighted in the revised manuscript.

Author response: Added reasoning for the lower scale and described how the elevation and aspect were calculated. Although the elevation data are available at a higher resolution, in our experiments, using the data at full resolution slightly decreased performance. This could be attributed to noise in elevation, or possibly slightly inaccurate coordinates of the weather stations in our dataset. The specific elevation and slope values are calculated via linear interpolation at the station coordinates.

8) It is not clear why the K-Nearest Neighbor method was chosen for spatial interpolation.

Author response:

The K-Nearest Neighbor approach was chosen because it provides a simple, non-parametric way to propagate station-level predictions in a physiography-conditioned feature space, requires no assumptions about spatial stationarity, and is computationally lightweight for operational deployment.

9) There are no results or references of publications on the conducted preliminary analysis with feature importance analysis based on CatBoost and SHAP values (lines 275-278).

Author response: The previous wording could be interpreted as if SHAP results were produced in this manuscript. We revised the sentence to clarify that the feature selection was guided by the SHAP/feature-importance analysis reported in our earlier work [15], but that we do not repeat the SHAP analysis and therefore do not present SHAP plots in the current manuscript. The revised text is highlighted in the updated version.

10) It is not clear for which time period the results presented in Table 1 and Figure 3 were obtained.

Author response: Added to text: For spatial model validation, we reserve 6 randomly chosen weather stations, while the remaining 30 stations are used for training. Temporally, the training data spans the years 2020–2022, while the validation data consists of measurements from 2023–2024.

11) It is not clear for which weather stations the results are presented in Figure 4.

Author response: The weather stations in Figure 4 (now Figure 6) are all of the 6 stations from the spatial validation set. That is mentioned in the description.

Author response: Changed case A figure to be from the temporal training subset instead. Added detailed description to both figures.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

After evaluating the revised version of the manuscript and the responses provided by the authors, I consider that the comments and observations made during the review have been addressed satisfactorily and substantively. The modifications strengthen the conceptual clarity, methodological transparency, analytical rigor, and discussion of results, improving the overall quality of the work. Therefore, I recommend acceptance of the manuscript.

There is only one minor detail related to the numbering of some citations and references (Line 108), which can be corrected routinely during the editorial process.

Author Response

Comments 1: There is only one minor detail related to the numbering of some citations and references (Line 108), which can be corrected routinely during the editorial process.

Response 1: Dear Reviewer,

Thank you for your positive evaluation of our manuscript and for pointing out the minor issue related to citation numbering in the Introduction (line 108).

Upon careful revision, we identified that the issue was not a general numbering problem, but an inconsistency between the content of the cited works and the statement they were supporting. In particular, references originally cited in this sentence referred to station-based time-series forecasting, while the sentence discussed machine-learning-based post-processing of numerical weather prediction (NWP) outputs.

We have corrected this by replacing the original citations with references that explicitly address ML post-processing of NWP forecasts. In addition, we reviewed the surrounding Introduction text and adjusted several citations to ensure that each reference is now used in a context that accurately reflects its thematic content.

These changes improve the conceptual consistency between the literature review and the reference list, without affecting the scientific results or conclusions of the manuscript.

Thank you again for your careful reading and helpful remark.
Authors

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This is my second round of review. Thank you for taking my comments into account. I still have some concerns with respect to the sections and the authors’ responses...

Author answered: “Author response: We agree. To explicitly validate the hypotheses, we added a dedicated subsection “4.5. Evaluation of hypotheses” in the Discussion, where H1–H5 are revisited and linked to the corresponding quantitative evidence (Tables 1–2) and relevant sections."

Please remove the use of “:” in the text and rewrite sentences. They should be written as a continuous textual description. For example, delete and modify line 738 and rewrite the paragraph as a textual analysis of the hypotheses. The same applies to lines 515, 536, 547, 566, 590, 693, and 731.

Finally, my previous concern was not addressed: are the data and code used in this proposal available? Can other researchers use them for the presented or new comparative analyses? Is the experiment fully reproducible?

Please explain the reproducibility aspects in the article.

In addition, I do not understand lines 353–354. The article states: “The overall architecture follows a modular design aligned with the ALIANCE [37], [38], [52], [53] platform…”, but it is not clear what workflow is being described. A figure similar to Figure 4 (which shows the infrastructure) should be added to explicitly illustrate the workflow itself.

Author Response

Comments 1: Author response: We agree. To explicitly validate the hypotheses, we added a dedicated subsection “4.5. Evaluation of hypotheses” in the Discussion, where H1–H3 are revisited and linked to the corresponding quantitative evidence (Tables 1–2 and relevant figures), with clear statements of the support level for each hypothesis. For clarity and consistency, the retained hypotheses are numbered sequentially as H1–H3 in the final manuscript.

However, subsection 4.5 (Evaluation of hypotheses) was not included in the article; instead, subsection 4.5 corresponds to “Strengths and limitations”. Therefore, hypotheses H1–H5 were not analyzed, and only a brief paragraph was included in Section 5.

Response 1: Thank you for noting this discrepancy. You are correct that, in the previous revised manuscript, the subsection titled “4.5. Evaluation of hypotheses” mentioned in our response was not present; instead, Section 4.5 corresponded to “Strengths and limitations,” and the hypotheses were only briefly summarized in the Conclusions.

We have corrected this in the revised manuscript as follows:

We added a dedicated subsection “4.5. Evaluation of hypotheses” within the Discussion, in which H1–H5 are explicitly revisited and linked to the corresponding evidence in Tables 1–3 and Figures 7–8, with a clear statement of the support level for each hypothesis.
The former Section 4.5 was renumbered to Section 4.6 (Strengths and limitations).
The brief “Hypotheses summary” previously appended to the Conclusions was removed and replaced with a cross-reference to Section 4.5 to avoid redundancy.

These changes ensure internal consistency between the manuscript and the response letter and provide the explicit hypothesis-level evaluation requested

Comments 2: Please remove the use of “:” in the text and rewrite sentences. They should be written as a continuous textual description. For example, delete and modify line 738 and rewrite the paragraph as a textual analysis of the hypotheses. The same applies to lines 515, 536, 547, 566, 590, 693, and 731.

Response 2:

We revised the indicated parts of the manuscript to eliminate “:” constructions and to replace colon-led list formulations with continuous prose. Specifically, we rewrote (i) the Research goals and hypotheses passage into a narrative description, (ii) the Static geographic data and Feature construction passages into single continuous sentences rather than list introductions, and (iii) the “Description and interpretation” paragraphs for Table/Figure results into standard continuous interpretation text. All edits are highlighted in yellow in the revised manuscript for straightforward verification.

Comments 3: Finally, my previous concern was not addressed: are the data and code used in this proposal available? Can other researchers use them for the presented or new comparative analyses? Is the experiment fully reproducible?

Response 3: We have addressed this by replacing the “Patents” section with a dedicated “Data and Software Availability” section. The manuscript now states that the primary input datasets used in the study are publicly available, including Global Forecast System predictors obtained from the National Center for Atmospheric Research Research Data Archive and the static physiographic layers derived from ASTER GDEM, while station observations are accessible through the SensLog platform. In addition, we clarify that the modelling workflow, feature construction, and evaluation protocol are described in sufficient detail in the main text to support independent re-implementation and comparative analyses.

At the same time, we explicitly state that our source code and operational data-processing pipeline are not publicly released, because the implementation is currently integrated into commercial applications.

Comments 4: Please explain the reproducibility aspects in the article.

Response 4: We have expanded the manuscript to explicitly address reproducibility in Section 6 (Data, Software, and Reproducibility). The added paragraph clarifies that (i) the primary input datasets are publicly accessible, and the predictor set, temporal coverage, spatial domain, and evaluation protocol are specified in sufficient detail to enable independent re-implementation and comparative analyses, while (ii) exact computational reproducibility through direct execution of the original operational pipeline is currently not possible because the production codebase and configuration are integrated into commercial applications and are therefore not publicly distributed. This addition makes the reproducibility scope explicit and distinguishes independent reproduction from rerunning the authors’ proprietary implementation

Comments 5: In addition, I do not understand lines 353–354. The article states: “The overall architecture follows a modular design aligned with the ALIANCE [37], [38], [52], [53] platform…”, but it is not clear what workflow is being described. A figure similar to Figure 4 (which shows the infrastructure) should be added to explicitly illustrate the workflow itself.

Response 5: We agree that the previous wording was ambiguous. We revised Section 2.3 to clarify that Figure 4 already represents both the modular architecture and the workflow, where the workflow is encoded by the directed dataflow between components (external data acquisition and ingestion, preprocessing and feature assembly, model training and evaluation, operational inference and super-resolution, and publication/integration via APIs). To make this explicit, we updated the text in Sections 2.3.1–2.3.2 and revised the Figure 4 caption accordingly.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

New version of the article contains a number of comprehensive corrections to the comments made earlier. We recommend this version of the article for publication in the MDPI Sensors journal

Author Response

Comments1: New version of the article contains a number of comprehensive corrections to the comments made earlier. We recommend this version of the article for publication in the MDPI Sensors journal

Response1: Thank you very much for your positive evaluation of the revised manuscript and for recommending it for publication in MDPI Sensors. We sincerely appreciate the time and effort you devoted to reviewing our work throughout the evaluation process.

Author Response File: Author Response.docx