This study presents a novel global 30 m resolution annual cropland dynamics dataset (GACED30) spanning 2000–2024, aiming to address the limitations of existing global cropland products in terms of temporal continuity, definitional consistency, and the monitoring of structural evolution. By constructing a gap-free SDC30 time series, introducing a spectral-semantic sample alignment strategy, and employing a CatBoost classifier combined with spatiotemporal post-processing, the authors generated the first annual, high-resolution cropland dynamics product with long-term continuity. The dataset demonstrates excellent performance in accuracy assessment and shows strong agreement with FAO national statistics, successfully revealing a “Global North-South divergence” pattern in cropland expansion and contraction. Overall, the methodology is innovative, the data product holds significant scientific value and application potential. Suggested to be published after revisions. The following are the questions and some mistakes in this manuscript:

• Consider adding a brief statement on data availability at the end of the abstract for completeness.
• Table 1 is comprehensive, but consider adding a column indicating the temporal coverage of each product to facilitate direct comparison with GACED30's 25-year record.
• The preliminary CatBoost model used for sample screening lacks detailed parameter specifications. Providing these details or including them in supplementary materials would enhance reproducibility. The CleanLab-based temporal cleaning process is mentioned but not quantitatively described. What confidence thresholds were used? How were temporal inconsistencies identified? These should be clarified.
• The selection of β = 0.2 ha/year as the threshold for significant change (within a 1 km grid) lacks explicit justification. Please provide reasoning for this threshold (e.g., based on expected noise levels or agronomic considerations).
• The comparison with other products (Table 2) is valuable, but some products (e.g., GLAD Cropland, ESA World Cover) have differing temporal baselines. The potential impact of temporal mismatches on comparative metrics should be explicitly addressed.
• Please further discuss based on the results of this study to enhance the depth of the discussion.
• The scale is missing in all the spatial maps. Some small maps lack the north pointer or latitude/longitude.
• Ensure consistency in formatting, particularly for datasets and software references (e.g., NASA JPL NASADEM citation).
• Please improve English expression in the manuscript.

The authors present the first global 30 m annual cropland extent dynamics product for 2000–2024, with a clear methodological framework, independent FAST-based validation, comparison against several global products, and consistency checks against FAOSTAT. The dataset itself is highly valuable and likely to be useful for global land-system research. That said, several issues should be addressed to further strengthen the manuscript.

Major comments

(1) The current validation relies mainly on the FAST validation set, comparisons with six global products, and consistency with FAOSTAT across 132 countries. This is useful, but still not enough to demonstrate accuracy in key agricultural regions, especially in areas with fragmented fields, complex cropping systems, or intensive land-use transitions. I strongly recommend adding comparisons with high-quality regional products and independent regional statistics, for example the CACD product in China and official cropland statistics such as the Third National Land Survey.

(2) The manuscript uses CatBoost as the core classifier and provides a brief rationale for this choice, but there is no direct comparison with other widely used machine-learning models such as Random Forest or XGBoost/LightGBM. Without such benchmarking, it is difficult to judge whether CatBoost is truly the most suitable model for this task. At minimum, the authors should include one or two baseline model comparisons and report both accuracy and computational efficiency.

(3) The independent validation appears to emphasize stable cropland, while transition areas are less well assessed. The manuscript states that the independent validation set was extracted from stable cropland records in FAST. This supports overall map accuracy, but it does not fully test the product’s ability to capture expansion, abandonment, active fallow, and other transitional states, which are central to the paper’s concept of “structural evolution.” A more targeted validation for transition pixels is needed.

(4) The discussion should expand the practical applications of the dataset. The manuscript already mentions implications for global sustainability, telecoupling, carbon accounting, and Earth system modelling, but these applications are still described rather generally. This section would be stronger if the authors more explicitly discussed how the dataset could be used in food security assessment, agricultural frontier monitoring, land-use policy evaluation, carbon accounting, and Earth system model benchmarking.

(5) The product is explicitly aligned with the FAO definition and includes permanent woody crops, active fallow, and agricultural structures, while excluding temporary meadows and pastures. This is reasonable, but it also means that part of the disagreement with other global products may be driven by definitional differences rather than mapping quality alone. The manuscript should make this point more explicit in both the validation and discussion sections.

Specific comments

(1) It would help to report performance separately by region or agro-ecological zone, rather than only globally.

(2) The discussion of potential applications should be more concrete. For example, please elaborate on use cases in food security, agricultural expansion monitoring, telecoupled land-system analysis, carbon accounting, and policy evaluation.

(3) Please clarify more explicitly that part of the mismatch with peer products may arise from different cropland definitions, especially regarding orchards, active fallow, and temporary meadows/pastures.

(4) The manuscript would benefit from a short paragraph stating where the dataset is expected to perform less well, for example in smallholder mosaics, highly fragmented systems, or regions with strong semantic ambiguity between cropland and grassland.

(5) Please check the data information. The pixel values in each band indicate land-cover classification status; however, I get quick look at the data, and it seems that cropland is coded as 10.

This study presents the annual updated cropland dataset at 30m resolution for the year 2000-2024. The proposed method integrates gap-free SDC30 time series with spectral-semantic sample alignment to generate such valuable dataset. Such datasets are essential for both crop monitoring community and land use land cover domains as it makes a substantial contribution to the Earth system science community by providing a temporally coherent, high-resolution dataset. While the manuscript is potentially publishable, there are still several major concerns, in particular the mixture of the different sub-classes of cropland. The samples used for training is binary cropland and non-cropland while the intra-class variations of different croplands (permanent crops, temporary crops, active fallow, etc) will hamper the classification accuracy. Meanwhile, in its current form, the manuscript suffers from significant shortcomings in methodological transparency, validation design, and demonstration of novelty relative to existing products. These deficiencies must be addressed before the dataset and its documentation meet the publication standards of Earth System Science Data.

Major Comments and suggestions:

1. The manuscript describes a spectral-semantic sample alignment strategy where candidate locations are stratified using the GLAD Cropland product and subsequently filtered by an expert-trained model. The final training set is, by design, heavily conditioned on the spatial priors of GLAD. This could artificially suppress the detection of cropland types that GLAD systematically misses (e.g., tree crops, permanent crops, etc). The authors must provide a rigorous analysis of the independence of the final training sample set from the GLAD prior. Specifically, what is the proportion of "augmented" samples that fell outside of GLAD cropland masks? Without this, the claim of outperforming GLAD in areas where GLAD is known to be weak is not fully substantiated. At the same time, I suggest to consider the dataset described in the peer-reviewed paper ‘A global reference dataset for land cover mapping at 10 m resolution’ to amend the training set.
2. The strength of the proposed dataset is the annually updated, inter-annual consistency. However, it is not clear how the proposed method ensure the consistency. Accordingly to the description in Line 227, the authors utilized year-specific CatBoost models to do binary classification. In this case, the CatBoost models trained in one year has no linkage with the next year or the year before. It is not convincing that such annual mapped cropland be consistency across years.
3. Inter-comparison methodology in particular for area estimation based on different datasets are inconsistent. In Section 3.5.2, the authors correctly apply sample-based area estimation only to GACED30. While the area totals and trends for GLAD, ESA WorldCover, and GLC_FCS30D are derived from simple pixel counting, the GACED30's area is adjusted for omission/commission errors. A pixel-count comparison between a bias-corrected map (GACED30) and uncorrected maps is not a fair assessment.
4. The paper explicitly excludes "Temporary meadows and pastures" to avoid overestimation but includes "Active Fallow." In reality, differentiating a fallow field (bare soil) from a plowed temporary pasture or a degraded rangeland using 30-m Landsat time series is extremely challenging spectrally.
5. Validation needs substantially enhanced. On one side, the validation shall targets not only for one year but also the 25 years. On the other side, validation over the inter-annual changes (expansion or reduction areas) shall be amended. While comparison with FAO statistical data provides quantitative accuracy metrics, it does not mean GACED30 outperformed other cropland layers or products are the metrics are a mixture of the classification accuracy and the discrepancy of definition of the classification schemes.

Other specific comments

1. In Table 1 - Temporal attributes for "ESA WorldCover" is "Annual 2000-2001". This appears to be a typo.
2. Line 315 (Section 3.4). The slope threshold of > 0.2 ha/year is defined. In a 1 km grid, 0.2 ha is roughly 2.2 Landsat pixels per year. Given that the product is 30-m resolution, is this threshold reasonable to define whether such changes are not "minor fluctuation"?
3. Figure 5 title states "GACED30 (2000-2022)" but the paper's title and dataset span 2000-2024. Please align the figure period with the full dataset period or explain why 2023-2024 is omitted from this specific visualization. Also, why the authors presented different conversion periods on the map of Brazil, China, and Kaz.
4. As presented in Figure 7 a, several non-agricultural areas or minor agriculture producing regions presented statistical significance of changing trends. Is this the consequences of classification bias or the actual phenomenon, for example, the significant changing trends in Sahara desert.
5. In Data Availability section, I expect the authors focusing on the products generated by this paper or the key inputs data produced by the team, while removing the information of other source of dataset. Or, you could document the way of data access for other land cover products in the Data part.