This study creates a consistent long-term and high-resolution global soil moisture product with comparable accuracy to previous deep learning approaches and demonstrates a clear superiority of the AtLSTM approach over other compared approaches. The findings and datasets are valuable to the community. The reporting is clear. Overall I recommend it for publication after minor revisions. Please see below for comments: 

1. The depth of the developed dataset should be noted in the abstract. 

2. The training target is a previously developed 1km dataset, and the newly developed dataset mainly provides the advantage of being longer-term and better winter coverage. This would be clearer if the temporal coverage of the Zhang et al. (2023) dataset is stated in p4 lines 117-118. 

3. p5 line 146: "8-day temporal resolution interpolated to daily" Please specify which interpolation method was applied. 

4. It would be nice to add 1-2 sentences about uncertainty in the sampling strategy of the training target (p9 lines 242-p10 ln 244). 

5. p13 Table 4: Should the number of layers be "2" for Bi-LSTM and 1 elsewhere? 

6. The sequence lengths in Table 4 span whole year. The testing on p17 only spans 0-29 days and demonstrates stabilized performances at much shorter lengths than 29 days. Why the drastic increase in sequence lengths in production runs? Also, since the input data are 3 discrete years, please specify what values are used to pad the 60 days before the start and after the end of the whole year. 

7. p14 lines 366-367: A bit confusing because this reduced-sample testing was not described in the Methods. Could you add a description? 

8. p15 Table 5: Is there a specific reason for choosing 0.4 m3/m3 as a threshold for large SM values?

9. Fig. 6: COSMOS dataset is not directly comparable to the developed dataset due to significant depth differences. It is okay as an auxiliary comparison, given that other comparable independent validation datasets are used (1982-1999 ISMN and SMAP validation sites). However, for informative purpose, it would be good to provide boxplots of the depths of the COSMO observations used in the last three columns of Fig. 6 in supplementary info. 

10. More on Fig. 6. The currently used 1982-1999 ISMN set provides spatiotemporally independent comparison. The 1982-1999 data at the 715 representative stations would still be temporally independent comparison. Could you include the performance metrics on this subset of data? 

This study used several deep learning models to estimate SM with some optical remote sensing indices, soil properties and DEM. Even though it’s a big work, it lacks sufficient innovation and employs inappropriate evaluation method. Despite the overall scale of the work, the methodological design raises serious concerns.

Major Comments:

Line 135, Table 2: I question the innovation and scientific contribution of using ERA5-Land SM as an input variable in a model that aims to estimate soil moisture. This approach may introduce circular reasoning and undermines the novelty of the proposed method.

Line 380, Table 5: It is inappropriate to use GLASS-MODIS SM as a reference to evaluate the model's estimated SM, as they are not independent. Since variables such as LST and albedo were also used in the model estimation, the evaluation becomes biased, leading to inflated performance metrics (e.g., R² > 0.98). This dependence compromises the reliability of the validation.

Line 20: The five deep learning models yielded very similar results, with R² values ranging from 0.982 to 0.987. Therefore, the claim that these models “…effectively enhance…” SM estimation is not well supported.

Minor Comments:

Line 60: There are additional remote sensing-based SM products that should be referenced, including but not limited to: Cheng et al. (2023); Guevara, Taufer, & Vargas (2021); and Zheng, Jia, & Zhao (2023).

Line 190: The time period referred to as "period1" is 2000–2018. Why were the years 2019–2021 excluded from the analysis?

Line 355: The statement “two widely used long-term global SM products” should specify which products are being referred to for clarity.

Cheng, F., Zhang, Z., Zhuang, H., Han, J., Luo, Y., Cao, J., . . . Tao, F. (2023). ChinaCropSM1 km: a fine 1 km daily soil moisture dataset for dryland wheat and maize across China during 1993–2018. Earth System Science Data, 15(1), 395-409. doi:10.5194/essd-15-395-2023

Guevara, M., Taufer, M., & Vargas, R. (2021). Gap-free global annual soil moisture: 15 km grids for 1991–2018. Earth System Science Data, 13(4), 1711-1735. doi:10.5194/essd-13-1711-2021

Zheng, C., Jia, L., & Zhao, T. (2023). A 21-year dataset (2000-2020) of gap-free global daily surface soil moisture at 1-km grid resolution. Sci Data, 10(1), 139. doi:10.1038/s41597-023-01991-w

The authors compared multiple deep learning methods and selected the most suitable approach to generate a 5 km-resolution soil moisture product based on AVHRR data, spanning four decades. Overall, this is a well-written manuscript, and the scope of the study aligns well with the aims and scope of Earth System Science Data. However, the current version of the manuscript appears overly technical in its presentation. The authors should address the following major and specific comments before the manuscript can be considered for publication.

General comments

1. While the description of the deep learning methods is sufficiently detailed, the treatment of the predictor variables requires substantial clarification and improvement. It remains unclear how these variables were selected, particularly the rationale for using ERA5-Land data as model inputs. In my view, the relatively strong performance of GLASS-AVHRR in capturing seasonal variations, especially for the R metric, may be partially attributed to this. The authors should consider producing an independently driven soil moisture product, rather than relying on predictors that may introduce redundancy or circular reasoning.
2. I remain concerned about the generalization capability of the developed deep learning model, particularly considering the strong spatial autocorrelation inherent in the current training, validation, and test data split. Furthermore, the results presented in Table 5 indicate that model performance deteriorates when soil moisture exceeds 0.4 m³/m³, which the authors attribute to a lack of training samples. This observation indirectly suggests that the model performs poorly in unseen regions or conditions, highlighting its limited generalization ability. The authors should explore alternative data splitting strategies to rigorously assess the robustness and generalizability of their approach.
3. Figure 2 in the Methods section is unnecessarily technical and lacks clear justification. A concise textual explanation that hyperparameter tuning was performed would be sufficient. Moreover, both Figures 1 and 2 fail to clearly illustrate how the spatiotemporal training was implemented, especially considering that LSTM and related models are specifically designed to capture temporal dependencies. The authors should provide a more explicit and structured explanation of how time-series characteristics were incorporated into the model training, without ignoring the spatial autocorrelation issue raised above.
4. In Figure 9, I noticed that the GLASS-AVHRR estimates provide coverage during the winter season, unlike the CCI product. The authors should clarify what measures were taken to ensure the reliability of these estimates under winter conditions, particularly in areas covered by snow, ice, or water. An additional seasonal evaluation focusing specifically on winter performance would be welcome and could strengthen the credibility of the product. In addition, all products in Figure 9 should be masked over permanent water bodies, where soil moisture retrieval is not valid. This issue is especially noticeable in Quebec, Canada.
5. The authors should provide more reasonable explanations for some of the evaluation results. For example, why does GLASS-AVHRR perform worse than ERA5-Land in terms of the R metric over COSMOS-Europe? In addition, I do not fully agree with the explanation provided for the strong dry bias observed in COSMOS-UK. ERA5-Land shows a consistently wet bias across different COSMOS networks, which raises the question of whether the developed product itself suffers from instability in different regions.

Specific comments:

L61: This may also be related to the uneven spatial distribution of the input meteorological data, particularly the limited observational coverage in tropical regions.

L64: also 10.1016/j.rse.2022.112921; 10.1016/j.rse.2022.113272

2 Datasets: Predictor variables lack of vegetation-related variables

Table 2: As noted above, it is inappropriate to use ERA5-Land-derived soil moisture as an input when the objective is to analyze or generate a new soil moisture product.

L146: What kind of method?

L156: see 10.1016/j.rse.2023.113721

L175: This also means that validation with ISMN is not independent.

L196: The meaning of the label "G" should be clearly defined.

L201: How do the authors justify the appropriateness of using COSMOS observations, which primarily sense deeper soil moisture, for evaluating surface soil moisture products?

Table 3: Are these COSMOS networks intercalibrated?

Sections 3.3 and 3.4 can be simplified

L377: The term "underestimate" should be used with caution, as all reference datasets represent proxy observations rather than true ground-truth measurements at the grid scale.

Figure 3: Lack of units

L506-507: Why doesn't ERA5-Land have this problem?

Figure 6: These three datasets, including ERA5-Land, GLASS-MODIS and GLDAS-AVHRR, have different spatial resolutions. Please clarify how the spatial mismatch was addressed when using different in-situ observations.

L529: Why doesn't ERA5-Land have this problem?

Figure 10: All products should be masked out over lake areas where soil moisture retrievals are not valid.

This paper introduces a dataset and approaches integrating AVHRR albedo and land surface temperature using an LSTM-based deep learning model to reconstruct soil moisture from 1982 to the present from GLASS-MODIS. The paper is well written, and the methodology is clear and well-detailed, but the paper does not highlight the novelty of the approach and how it differentiates from other efforts in the field. One major concern is that GLASS-MODIS already uses ERA5-Land, MERIT-DEM, and Soil-Grids as inputs; as such the proposed DL models are simply acting as surrogate models to GLASS-MODIS. The authors should demonstrate that AVHRR and the proposed DL model substantially add value to the predictions of soil moisture. As it reads, the added value of the approach is limited to extending the soil moisture time series to the AVRHH time series, which is helpful to evaluate long-term trends, but I wonder how it adds value beyond a CDF matching bias correction, for example.

Major:

GLASS-MODIS already uses ERA5-Land, MERIT-DEM, and Soil-Grids, which are also input data for your approach in this paper. How do you ensure your DL model is actually learning anything from GLASS-AVHRR and not just a surrogate for GLASS-MODIS learning directly the relationship from ERA5-Land, MERIT-DEM and SoilGrids. In other words, the if the target data is not independent from the input data, your DL model is just a surrogate for the existing GLASS-MODIS model. Two ways to quantitatively test for that is to 1) check how statistically significant is the difference in the accuracy of your trained DL models with respect to GLASS-MODIS when validating the approach with in-situ observations, and 2) by testing how different is your DL accuracy if trained only on ERA5-Land, MERIT-DEM and Soil-Grids for the past period (1982-1999). These experiments would help elucidate how the proposed model is actually learning new things from the AVHRR data or just mimicking GLASS-MODIS from ERA5-Land, MERIT-DEM, and Soil-Grids.

Moderate:

R2 0.987 indicates that probably the DL is overfitting or there is contamination between the training and testing datasets used. In fact, when using the model to produce the GLASS-AVHRR dataset accuracy drops from R 0.63 to 0.79

Minor:

I appreciate how the authors ensured that testing grids are within 25km distance from training grids. This prevents the testing sample is contaminated with inputs from the training datasets (since the resolution of precipitation input data from ERA5 into ERA5-Land runs is about 25km resolution).

Please add GLASS-MODIS to Figure 6 for the COSMOS comparisons. Please include marks indicating whether the difference between the accuracy of the SM data products is statistically significant.

Please add GLASS-MODIS to Table 6, which will be best displayed in landscape format. Also, please mark in bold the best statistically significant data product/metric for each site.

Please add GLASS-MODIS to Figure 7.