The authors pursue the ambitious objective of developing a global, high-resolution GWSA dataset at a spatial resolution of 1 km. Although the resulting data set could provide a useful resource for groundwater research, the manuscript does not adequately demonstrate its reliability. My major comments are as follows.

(1) As shown in Table 1, TWSA data extend through the end of 2025, whereas the other input data sets end in 2020. This five-year temporal mismatch requires justification. Why was the GWSA data set not updated through 2025? Other satellite-based products that are regularly updated and currently available should be considered to extend the temporal coverage of the input data.

(2) The validation is insufficient to establish the reliability of the proposed data set. First, the estimated GWSA should be evaluated against available groundwater-level observations from monitoring wells. Second, the authors should select representative groundwater basins and compare their GWSA estimates with those reported in previous regional studies, with particular attention to differences in temporal variability, magnitude, and long-term trends. Third, comparisons with simulations from existing continental- or global-scale land surface and groundwater models would provide an additional independent assessment. Validation should not be limited to correlation analysis against an existing GWSA product. For comparisons with monitoring-well observations, the authors should also present and interpret time-series plots to determine whether the estimated GWSA captures observed seasonal fluctuations, interannual variability, and long-term trends.

(3) Products derived from different data sources should be systematically compared and cross-validated. The manuscript should quantify the consistency and discrepancies among these products and discuss how differences in data sources propagate into the final GWSA estimates.

(4) As shown in Table 2, assigning a uniform specific yield value to each of the seven regions may not adequately represent the spatial heterogeneity of aquifer properties. Because GWSA estimates derived from groundwater-level observations are sensitive to Sy, the authors should evaluate the sensitivity of their results to the selected Sy values and discuss the uncertainty introduced by this regional parameterization.

(5) The manuscript should provide a more detailed description of how temporal data gaps in the GRACE and GRACE-FO records were handled. The authors should specify the gap-filling method, its underlying assumptions, and the uncertainty it introduces into the reconstructed GWSA time series.

(6) Each abbreviation should be defined only at its first occurrence and used consistently thereafter. Repeatedly spelling out a term after its abbreviation has already been defined is unnecessary.

I think the 1-km global groundwater storage anomaly (GWSA) dataset downscaled from GRACE satellite observations presented in this manuscript could be valuable for hydrological research and water-resource assessment. However, I have several substantial concerns, particularly regarding the physical definition of the GWSA target and the temporal consistency of the GLDAS inputs. Addressing the issues below would substantially improve the physical consistency, transparency, and interpretability of the dataset.

Major comments

1. For Eq. (15), the manuscript states that “Terrestrial Water Storage (TWS) represents the aggregate water reserves both above and beneath the land surface, encompassing surface water, soil moisture, groundwater, snow cover, ice, and vegetation water storage.” However, Eq. (15) does not explicitly account for surface-water or ice-storage anomalies. Although glacier/ice storage may reasonably be neglected in typical non-glaciated land areas, surface-water storage (including rivers, lakes, reservoirs, wetlands, and floodplains) cannot simply be ignored in a global-scale analysis. I am not convinced that applying a permanent-waterbody mask adequately resolves this issue. First, masking removes spatial pixels, whereas subtraction separates mass-storage components. Masking water-body pixels does not eliminate the gravity signal associated with temporal changes in lake or river storage, including potential leakage into adjacent land areas. Second, many important surface-water variations are associated with non-permanent water bodies, including seasonal floodplains, temporary inundation, reservoir fluctuations, and wetland-storage changes. Consequently, if surface-water storage increases but is not explicitly removed from TWSA, the inferred GWSA may be overestimated; conversely, decreases in surface-water storage could potentially be misinterpreted as groundwater depletion. I therefore suggest that the authors more explicitly discuss and, if possible, quantify the uncertainty introduced by the omission of dynamic surface-water storage. This issue may also be relevant to the interpretation of Fig. 5. In particular, could the dendritic pattern resembling a river network in the Amazon Basin partly result from residual surface-water signals that were not explicitly removed, rather than representing groundwater-storage variability alone? I believe this alternative explanation deserves further examination.

1. In line 165, the manuscript refers to “the annual time-series data of SWS, SWE, and CWS”. Is “annual” a typo, or were annually aggregated data actually used rather than monthly data? If annual values were indeed used together with monthly GRACE TWSA, this would introduce an important temporal-scale inconsistency. Monthly variations in soil moisture, snow water equivalent, and canopy water storage that should have been removed from TWSA would remain embedded in the inferred groundwater component. This could be particularly important in monsoon regions, snow-dominated regions, and other areas with strong seasonal land-water-storage variability. If so, the conclusion in lines 698-699 that the seasonal contribution exceeds 50% across 52.3% of the global land area would also require reconsideration, because part of the seasonal signal attributed to groundwater could instead originate from seasonal variations in soil moisture, snow, or canopy storage that were not removed at the appropriate temporal resolution. If “annual” is simply a wording error and monthly GLDAS data were actually used, this should be corrected explicitly. In either case, it would also be helpful for the authors to provide the exact GLDAS-Noah variables used, their units, the treatment of the soil-moisture layers, and the temporal aggregation and anomaly-calculation procedures.

Minor comments

Line 337: Please clarify what is meant by the “Warp function”.

Lines 368-370: The study uses a random 70%/30% split for model training and validation. Because groundwater and hydroclimatic datasets generally exhibit substantial spatial and temporal autocorrelation, a random split may lead to overly optimistic estimates of model performance if nearby pixels or temporally adjacent observations occur in both the training and validation sets. I recommend supplementing the current validation with more stringent approaches, such as spatially blocked and temporally blocked validation.

Line 425: I agree with the comment from the first referee regarding the specific-yield data. Given that SY can exhibit pronounced spatial heterogeneity, the use of broad regional averages may obscure important hydrogeological differences. It would be useful to discuss whether higher-resolution SY information could improve the in-situ validation.

Lines 509-510: Statements such as “the sharp boundaries caused by localized human pumping” should, in my view, be expressed more cautiously. The additional fine-scale spatial information in the 1-km product necessarily originates largely from the high-resolution predictor variables, because GRACE itself does not resolve groundwater-storage variability at 1-km scales. Therefore, the presence of sharp spatial boundaries in the downscaled product does not by itself demonstrate that these boundaries are directly caused by localized groundwater pumping. Independent high-resolution observations would be needed to support such an interpretation.

Figure 4: The use of 1,518 monitoring wells is potentially an important strength of this study. However, the validation procedure involves several levels of aggregation, including aggregation into 15 bins. Such binning suppresses local variability and observational noise and may therefore increase the apparent agreement between modeled and observed trends. I recommend additionally reporting the unbinned station- or grid-level validation results, including R², RMSE, bias, and the corresponding scatter plot. A map showing the spatial pattern of the in-situ observations would also be helpful for assessing the geographical representativeness of the validation dataset.

Section 3.5: The interpretation of the residuals as anthropogenic pumping signals should be more conservative. The inferred anthropogenic component is obtained by subtracting GHM-simulated background signals from the reconstructed GWSA. As the manuscript itself acknowledges, these residuals inevitably include structural errors from the GHMs, in addition to uncertainties in the GRACE-derived GWSA and the downscaling procedure. I therefore suggest referring to these residuals as signals consistent with anthropogenic groundwater depletion, or as depletion signals not explained by the modeled natural background, rather than treating them as direct estimates of groundwater pumping.