Given that near-surface humidity and temperature are critical drivers of glacier ablation (surface energy balance), permafrost degradation, and alpine ecosystem evolution, this manuscript addresses the severe scarcity of centennial-scale, high-resolution datasets over the Tibetan Plateau by proposing an innovative hybrid-structure deep learning downscaling approach based on FourCastNet. Nevertheless, the current manuscript still exhibits several critical limitations that must be addressed through a major revision before it can be considered for publication.

Major

1. The paper utilizes FourCastNet for data reconstruction, but the manuscript currently lacks details regarding the hyperparameter tuning process and specific parameter configurations. For an ESSD manuscript, reproducibility is a critical criterion. The authors are strongly encouraged to provide a comprehensive hyperparameter table and specify the training environment details in the main text or appendix.
2. The authors utilized standard deterministic metrics (R², RMSE, MAE, and IOA) to quantify reconstruction errors. However, this approach treats meteorological station observations and the TPMFD target dataset as the absolute "ground truth," neglecting the inherent instrumental errors and the systematic uncertainties embedded within the forcing products themselves. To enhance the physical reliability of TPHH for cryospheric and hydrological applications, the authors should account for these error sources and provide spatiotemporal uncertainty estimatesfor the reconstructed dataset.
3. The description of the different station numbers (87, 125, 135) is very confusing to follow, and their respective roles remain unclear. In particular, it is difficult to assess whether the validation stations are fully independent from the TPMFD target dataset, which itself incorporates CMA station observations. The authors should explicitly clarify:1) which stations were used in TPMFD construction; 2) which stations were excluded from training; 3) which stations were used for truly independent evaluation,

and 4) Whether there is any duplication of data information between the target dataset and the validation procedure. In addition, Figure 7 appears to use the 125/135 CMA stations rather than the 87 independent stations mentioned earlier in the manuscript. The rationale for this choice is unclear. If independent stations are available, it would seem more appropriate to use them for the primary validation analyses in Figure 7 to avoid potential dependence between TPMFD and the validation dataset.

4) The reconstructed 1901–1978 historical climate fields rely entirely on a deep learning model trained on the 1979–2023 modern period. This implicitly assumes statistical stationarity in the relationship between the coarse CRU predictors and the fine-scale TPMFD targets across vastly different climatic regimes. However, this critical assumption is not rigorously evaluated. Prior to 1979, station coverage over the Tibetan Plateau was extremely sparse, and early CRU grids heavily rely on large-scale interpolations that lack real mesoscale gradients. Applying a model overly optimized for modern terrain features to these early interpolated fields risks generating artificial spatial features or "homogenized" artifacts. The brief mention in Section 6.3 is insufficient. The authors must: (1) refrain from treating the reliability of the early-century reconstruction on par with the post-1979 period, and (2) substantially expand the discussion on how sparse early observations may affect the robustness, spatial variability, and uncertainty of the reconstructed 1901–1978 climate fields.

1. The spatial representativeness of the existing 135 validation stations exhibits significant limitations (Figure 1). Most stations are clustered in the eastern and southern Tibetan Plateau, particularly within relatively low-elevation valleys, while the central high-elevation interior and western Tibetan Plateau remain sparsely monitored or largely unrepresented. As a result, the current validation framework may not adequately assess model performance over the full Tibetan Plateau domain, especially in data-scarce high-altitude regions where climatic and topographic conditions differ substantially from the station-dense areas.
2. For a century-long climate dataset, the representation of long-term climate variability and trends carries the highest scientific value. The authors are recommended to provide spatial distribution maps of the long-term trends for near-surface humidity. Furthermore, these trends should be horizontally compared with existing centennial historical datasets (e.g., 20CRv3, the ERA5-Land back-extension, or older statistical downscaling products) to demonstrate the robustness of TPHH in capturing large-scale climate wetting/drying signals over the past 120 years.
3. The rationale for selecting the single-point validation location (91.1°E, 31.5°N) in Figure 11 is unclear. Since climatic variability over the Tibetan Plateau is highly heterogeneous, a single-point example is insufficient to demonstrate temporal continuity or reconstruction robustness over the entire domain. The selected point may also represent a favorable region with relatively smooth terrain and strong model performance. It is necessary to conduct further analysis using multiple representative sites from different climatic and topographical regions on the Tibetan Plateau, so as to more reliably assess the temporal continuity and reliability of the reconstruction results.
4. As a manuscript centered around datasets and reconstruction methods, scientific reproducibility is paramount. However, the current "Data availability" section of the manuscript only provides download link for the dataset, without mentioning any public addresses for the core code of the deep learning model. It is strongly recommended that the authors follow the open science policy of ESSD and provide publicly accessible DOI links.

Minor:

1.The major acronyms should be fully expanded upon their first appearance in both the Abstract and the main text. The acronyms like "CRU" and "CMA" are not spelled out at their first mention. Please check the entire text for similar errors.

2.Line 70, CRU or CRU_TS?

1. To help readers visually evaluate the spatial independence of the validation network, the authors are strongly encouraged to upgrade Figure 1 by:1)Utilizing distinct colors or geometric shapes to differentiate the historical stations prior to 1979 from the modern ones; 2)Clearly marking the truly independent validation stations; 3)Highlighting the location of the single-point validation grid used in Figure 11 (91.1°E, 31.5°N);
2. Despite repeated discussions throughout the manuscript regarding the profound impacts of terrain effects on downscaling meteorological fields over the rugged Tibetan Plateau, static terrain channels such as Digital Elevation Model (DEM), slope, or aspect are not explicitly included in the model's actual input variables (predictors). Why is this?
3. The manuscript frequently describes the TPMFD dataset as the "ground truth". In geoscientific and remote sensing literature, "ground truth" is strictly reserved for direct, in-situ physical measurements (e.g., weather station or flux tower observations). Since TPMFD is itself a blended/assimilated forcing product rather than primary direct observations, calling it "ground truth" is scientifically inaccurate. To maintain academic rigor, the authors should replace this term with “reference dataset” or “target dataset” throughout the text.
4. In Section 2 ("Comparable datasets"), the authors could explicitly provide the specific temporal coverage/time range for each climate product mentioned in the manuscript.
5. Line 353: ERA5 (0.1°)-> ERA5 (0.1°×0.1°)
6. The font of "temp, (b) shum, and (c) pres" in the caption of Figure 11 is inconsistent with the rest of the caption.

General comment

This study addresses a critical challenge in high-altitude climatology by constructing TPHH, a long-term (1901–2023), high-resolution (1/30°) near-surface humidity, temperature, and surface pressure dataset for the Tibetan Plateau. The work is highly significant as it bridges the pre-satellite era data gap, providing much-needed spatial detail for a region characterized by extreme topographic complexity and elevation-dependent warming. The methodological innovation lies in the application of the FourCastNet architecture, which synergistically integrates Vision Transformers and Adaptive Fourier Neural Operators to map coarse-resolution signals from the CRU dataset onto a high-resolution grid governed by terrain constraints. The validation effort is commendable, employing not only a network of 135 independent meteorological stations but also an innovative biological cross-verification using 41 tree-ring proxy groups to ensure temporal fidelity in the early 20th century. This dataset could provide a foundational material for studying long-term hydrological cycles, glacier dynamics, and ecosystem responses to climate change on the Tibetan Plateau. To meet the stringent standards of ESSD, it is essential that the authors thoroughly address the concerns raised below, specifically regarding the physical grounding of the FourCastNet architecture and the reliability of reconstructions in the data-sparse western Tibetan Plateau.

Major comments

1. Quantification of uncertainty in the pre-1950s observational void

The authors utilize CRU TS 4.08 as the primary predictor. It is well-documented that before the 1950s, ground-based observations within the Tibetan Plateau were virtually non-existent, and the CRU signal in this period relies heavily on spatial interpolation from distant stations outside the plateau. While the deep learning model successfully learns fine-scale "terrain textures" from the TPMFD training period (1979–2023), there is a risk that the reconstructed high-resolution features in the early 20th century are merely reflections of static topographic constraints rather than actual climatic fluctuations. The authors must provide a deeper discussion or a sensitivity analysis comparing the model's error characteristics before and after the densification of the observation network (e.g., around 1950).

2. Theoretical justification of physical consistency

The manuscript repeatedly emphasizes that the hybrid-structure deep learning approach is physically-consistent. However, FourCastNet is essentially a data-driven super-resolution model. While it effectively captures "terrain governance," it is unclear whether the model architecture or loss functions explicitly enforce thermodynamic laws (e.g., the Clausius-Clapeyron equation or the relationship between air pressure and hypsometry). The authors should clarify if this consistency is merely a spatial correlation with topography or if the model satisfies fundamental meteorological equations during the downscaling process.

3. Spatial Bias in Validation and Representative Uncertainty

As shown in Figure 1 and Figure 7, the 135 validation stations are predominantly clustered in the eastern and southern edges or river valleys of the plateau. The western interior and the high-altitude “No Man's Land” (Qiangtang) are almost entirely unrepresented in the ground-truth network. Given that the model's RMSE is higher in complex terrain, the authors should provide a spatial uncertainty map or an analysis of residuals across elevation gradients to quantify the reliability of the data in these unmonitored high-altitude regions.

4. Phenological logic of tree-ring variable selection (equation 5)

In the ternary regression model for tree-ring validation, the authors selected April-May-June (AMJ) humidity as a key factor. The choice of this specific seasonal window requires stronger phenological justification. In many parts of the Tibetan Plateau, tree growth may be more sensitive to JJA (Summer) humidity or have lagged responses to winter moisture. The authors should explain why the AMJ window was selected as the primary moisture constraint across all 41 proxy groups.

Minor Comments

Line 22 & Supplement (Fig S1, S2 titles): There is an inconsistency in the dataset’s acronym. The supplement uses “THPP” while the main text uses "TPHH". Please standardize to TPHH throughout.

Line 34: The text mentions “87 independent ground-based meteorological stations”, but Line 56 and Figure 1 refer to “135 stations”. Please verify the exact number used for historical validation (1901–1978).

Line 225 (Equation 1): The formatting for the R2 formula is slightly garbled in the PDF, with overlapping characters in the denominator. Please re-typeset using standard LaTeX.

Line 290: The text states, “As demonstrated in Table 1...”, but the performance metrics are actually presented in Table 3. This is a citation error.

Lines 505, 520, 530 of References: Several citations are listed for 2025 or 2026 (e.g., Jiang et al., 2025; O’Neill et al., 2025). Please indicate if these are preprints or currently in press.

Figure 6 (RMSE Heatmap): The color scale for specific humidity is quite narrow (max 0.40 g/kg). It is difficult to discern spatial variance in the plateau interior. Consider using a more contrast-enhancing color ramp for the humidity panel.

Figure 12: The Pearson correlation coefficient (0.82) is unexpected high. Please include the significance level (p-value) to strengthen the statistical argument.