Annual 10-m high-resolution cropland maps for Southeast Asia since 2019 using AlphaEarth embeddings

Cai, Yaotong; Li, Baoni; Liu, Xiaoye; Jiang, Xin; Zhu, Yakun; Luo, Shuxin; Qin, Yingzuo; Xie, Shudi; Ye, Jianhuai; Shen, Huizhong; Guo, Zhilin; Liu, Xiaoping; Zeng, Zhenzhong

doi:10.5194/essd-2026-78

Preprints

https://doi.org/10.5194/essd-2026-78

Preprints

25 Feb 2026

| 25 Feb 2026

Status: this preprint is currently under review for the journal ESSD.

Annual 10-m high-resolution cropland maps for Southeast Asia since 2019 using AlphaEarth embeddings

Yaotong Cai, Baoni Li, Xiaoye Liu, Xin Jiang, Yakun Zhu, Shuxin Luo, Yingzuo Qin, Shudi Xie, Jianhuai Ye, Huizhong Shen, Zhilin Guo, Xiaoping Liu, and Zhenzhong Zeng

Abstract. Southeast Asia (SEA) contributes substantially to tropical agriculture, but remains underserved by high-precision cropland data due to persistent cloud cover, fragmented farming, prevalent shifting cultivation, and complex phenology. Here, we developed a 10-m annual cropland dataset for SEA (SEA_Cropland10) covering 2019–2024 using a random forest model on Google Earth Engine. The model integrated 88,088 Sentinel-1 SAR scenes, 599,255 Sentinel-2 optical images, and Google AlphaEarth embeddings, and was trained using 37,192 visually interpreted samples. An independent accuracy assessment was performed using 1,200 samples stratified by land-cover change trajectories to validate both temporal dynamics and spatial extent. SEA_Cropland10 achieved an overall accuracy (OA) of 92.67 % (±1.47 %) for cropland dynamics, with annual static accuracies consistently exceeding 92.42 % (±1.50 %). The incorporation of AlphaEarth embeddings proved critical, improving model performance by 5.81 %. Compared to existing global products (e.g., GLAD, WorldCereal), SEA_Cropland10 improved OA by 14.56 %–19.97 % and substantially enhanced the detection of sloping cropland (>5°) in mountainous regions of SEA, increasing producer’s accuracy in global baselines from below 29.7 %–32.9 % to 89.0 %–99.1 %. Consequently, we identified three- to fourfold more sloping cropland area than GLAD and WorldCereal reported. Based on SEA_Cropland10, the estimated total cropland area in SEA shifted from 68.7 (±2.8) Mha in 2019 to 67.3 (±3.2) Mha in 2024, showing strong consistency with national statistics (r = 0.90–0.95). This dataset provides an important improvement for regional food security monitoring and carbon cycle modeling. The SEA-Cropland10 is publicly available at Zenodo: https://doi.org/10.5281/zenodo.17828801 (Cai and Zeng, 2026) and Google Earth Engine App: https://ee-caiyt33tc.projects.earthengine.app/view/seacrop10.

Received: 28 Jan 2026 – Discussion started: 25 Feb 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 1961 KB)

Supplement (172 KB)

Download & links

Yaotong Cai, Baoni Li, Xiaoye Liu, Xin Jiang, Yakun Zhu, Shuxin Luo, Yingzuo Qin, Shudi Xie, Jianhuai Ye, Huizhong Shen, Zhilin Guo, Xiaoping Liu, and Zhenzhong Zeng

Status: open (extended)

Post a comment Subscribe to comment alert

RC1:
'Comment on essd-2026-78', Anonymous Referee #1, 26 Aug 2026 reply
This manuscript presents an annual 10-m active-cropland dataset for Southeast Asia for 2019-2024, produced by combining Sentinel-1, Sentinel-2, terrain variables, and AlphaEarth embeddings in a random-forest framework. The regional focus, the attempt to map fragmented and sloping agricultural systems, and the release of a multi-year product are potentially valuable. Nevertheless, the manuscript does not yet provide sufficiently rigorous evidence for its central claims of extremely high model accuracy, improved detection of sloping cropland, and substantially larger mapped cropland area than some existing products.
The principal problem is the accuracy assessment. The validation design, splitting strategy, weighting scheme, and sample composition are not reported in sufficient detail to support the manuscript's quantitative claims. Additional concerns include poorly characterized cropland and non-cropland reference classes, inadequate assessment of boundary errors and their effect on area estimates, insufficient evidence that AlphaEarth specifically improves sloping-cropland detection, uncertain temporal transfer from a model trained only in 2024, and comparisons among products with different definitions and spatial supports. These issues require substantial new analyses rather than editorial clarification. I therefore recommend major revision.
Major comments
Model accuracy assessment. The validation design is not sufficiently transparent or statistically convincing. The manuscript states that five-fold cross-validation was used, but it does not report the splitting strategy. If the folds were generated by random point splitting, the result is not a valid test of map generalization because nearby pixels/patches, repeated observations from the same landscape and texture, or temporally correlated samples can occur in both training and validation folds. Such spatial and temporal autocorrelation can strongly inflate accuracy in land-cover mapping. The authors should replace or supplement random cross-validation with spatial block cross-validation, country- or eco-region-holdout tests, and temporally independent testing, and they should state exactly how all folds were constructed.

The cropland and non-cropland reference classes (in training and testing datasets) are not described in enough detail. The model-development dataset contains 22,821 cropland and 14,371 non-cropland samples from 2024, but the manuscript does not report the composition of either class in sufficient detail for training or testing. This is a serious problem because the difficulty of the classification depends strongly on what the negative class contains. If many non-cropland samples are bare soil, water, impervious surface, or other easily separable classes, the task is much simpler than if the negative class is dominated by vegetations like forests, shrublands, grasslands, fallow fields, young plantations, recently cleared land, wetlands, harvested bare soil, or shifting-cultivation mosaics. Similarly, the cropland class may include different crop species, field sizes, phenologies, management systems, and topographic settings, but the manuscript does not report whether these groups are represented or whether accuracy differs among them. The authors must report sample numbers by crop type, non-cropland subtype, country, slope, field size, and landscape context; provide confusion matrices for the full test set and all subgroups of crops and non-crops; and disclose the reference imagery dates, interpretation unit, ambiguity rules, interpreter independence and blinding, and inter-interpreter agreement. Because the independent test set is not very large, the authors should also provide representative RGB or Sentinel composite images with reference labels and predicted labels for cropland and non-cropland samples, and ideally archive the complete sample-level table so readers can examine the labels and errors.

Cropland boundaries and resulting area estimates are not adequately validated. By visual comparison between SEA_Cropland10 and other products in the GEE app provided by authors, it seems that products often agree in field interiors but differ markedly at field boundaries. This is a central accuracy issue in Southeast Asia, where many fields are small and fragmented and boundary pixels can represent a large proportion of the mapped cropland area. A systematic outward or inward displacement of only one 10-m pixel can substantially alter regional area estimates even when point-based overall accuracy remains high. The reported pattern of very high producer's accuracy but lower user's accuracy for sloping cropland is also consistent with a more expansive classification that reduces omission while increasing commission. Consequently, the finding that SEA_Cropland10 maps with several times more sloping cropland than GLAD or WorldCereal cannot presently be interpreted as evidence that the new product has recovered genuine omissions. It may partly reflect boundary expansion, mixed pixels, post-processing, and the comparison of binary products at different resolutions. The authors should add independent field-polygon validation, distinguish field interiors from edge zones, report object- or boundary-based measures such as IoU or boundary-F1, quantify area bias by field size, and test how one-pixel erosion and dilation affect national and regional cropland estimates. And the author should also define by how many percent of pixels are included in the polygon can be considered as a “True” pixel. Product comparisons should also be repeated after aggregation to a common grid using fractional cropland cover, with error-adjusted area estimates and confidence intervals.

The reported independent accuracy also appears to be unweighted under a disproportionate stratified design. The 1,200 independent validation sites were deliberately allocated among five mapped strata with very different sample sizes and mapped areas, yet the equations and reported metrics appear to be based on an unweighted sample confusion matrix. In particular, the dynamic overall accuracy of 92.67% can be reproduced exactly from the reported stratum sample sizes and user's accuracies: [300(0.9633) + 200(0.8700) + 500(0.9920) + 100(0.7400) + 100(0.7900)]/1200 = 0.9267. This is a sample accuracy, not a design-based regional accuracy, because the five strata do not occupy equal proportions of the mapped region. The annual overall, producer's, and user's accuracies, their confidence intervals, and the comparisons with GLAD and WorldCereal may therefore be statistically incorrect. The authors should provide the mapped area and inclusion probability of every stratum, reconstruct area-weighted error matrices, and recalculate all accuracy measures and paired product differences with design-based confidence intervals. The full validation table, including strata, annual reference labels, and map classes, should be archived so that the estimates can be independently reproduced.

The claimed AlphaEarth contribution to sloping-cropland detection is not demonstrated. Lines 20-25 of the abstract state that AlphaEarth embeddings proved critical, improved model performance by 5.81%, and substantially enhanced detection of sloping cropland in mountainous Southeast Asia compared with GLAD and WorldCereal. This causal interpretation is not supported by the current evidence. A comparison between the final product and existing global products cannot show that AlphaEarth, rather than training labels, Sentinel-1, Sentinel-2, terrain variables, post-processing, class definitions, or map resolution, caused the apparent improvement. The authors should provide targeted ablation experiments for sloping cropland, including models with and without AlphaEarth under the same samples, features, hyperparameter search, and validation design. They should report producer's and user's accuracy, commission and omission errors, and area bias for sloping cropland separately, with design-based confidence intervals. If the gain mainly occurs in field interiors but not boundaries, or if user's accuracy declines while producer's accuracy increases, the interpretation should be revised. The statement that AlphaEarth enables superior detection of mountain cropland should be weakened unless this mechanism is demonstrated directly.

Temporal transferability, product comparisons, and reproducibility remain insufficient. A classifier trained exclusively with 2024 labels is applied retrospectively to 2019-2023, yet the manuscript does not test whether its errors remain stable through time. The mapped regional area decreases from 69.4 Mha in 2019 to 59.4 Mha in 2022 and then rebounds to 67.7 Mha in 2024; this synchronized decline may reflect real agricultural change, but it also coincides with the loss of Sentinel-1B and may be affected by annual variation in cloud-free Sentinel-2 observations. The authors should provide annual valid-observation layers, assess the Sentinel-1B discontinuity, and report accuracy before and after temporal and morphological post-processing. The comparisons with GLAD, WorldCereal, GCEP30, and FAOSTAT are also not sufficiently harmonized because the products differ in class definitions, fallow treatment, temporal windows, thresholds, resolution, and grid support. Paired product comparisons should use the same reference definition and design-weighted uncertainty. Finally, the workflow is not reproducible enough: the full 187-feature stack, selected 83 features, compositing periods, Whittaker parameters, missing-data handling, resampling rules, random seeds, final random-forest implementation, parameter translation between scikit-learn-style settings and Google Earth Engine Smile Random Forest, SHAP and feature-selection code, validation labels, and post-processing settings should be released.

Impact of satellite image quality on the final product. The author should explain how the clouds, cloud shadows or other quality issues of satellite images affect the quality of final product. As this study uses a pixel-based model (random forest), if a pixel is affected by cloud, then the prediction should not be correct. In Asia-Pacific regions there are several months with heavy rainfalls or storms, the cloud can be a big issue. This will not be present in current validation strategy, as no cloud images are retained for testing, the author must include such images in testing dataset to assess the real uncertainties of the final products.

Minor comments
The title should state 2019-2024 rather than 'since 2019' unless an operational update plan is provided. Percentage-point differences should be distinguished from relative percentage improvements; for example, the difference between 95.79% and 90.53% is 5.26 percentage points, although the relative increase is 5.81%. Evaluative terms such as 'breakthrough', 'remarkable', 'exceptional', and 'massive' should be replaced by neutral quantitative language. Every reported '+/-' value should be identified explicitly as a standard error, confidence interval, or variation among cross-validation folds. The product name should be standardized as either SEA_Cropland10 or SEA-Cropland10, and the figure-panel references should be checked throughout.

Reply
Citation: https://doi.org/10.5194/essd-2026-78-RC1

Yaotong Cai, Baoni Li, Xiaoye Liu, Xin Jiang, Yakun Zhu, Shuxin Luo, Yingzuo Qin, Shudi Xie, Jianhuai Ye, Huizhong Shen, Zhilin Guo, Xiaoping Liu, and Zhenzhong Zeng

Supplement

https://doi.org/10.5194/essd-2026-78-supplement

Data sets

SEA_Cropland10: Annual 10-m high-resolution cropland maps for Southeast Asia since 2019 Yaotong Cai and Zhenzhong Zeng https://doi.org/10.5281/zenodo.17828801

Yaotong Cai, Baoni Li, Xiaoye Liu, Xin Jiang, Yakun Zhu, Shuxin Luo, Yingzuo Qin, Shudi Xie, Jianhuai Ye, Huizhong Shen, Zhilin Guo, Xiaoping Liu, and Zhenzhong Zeng

Viewed

Total article views: 1,597 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
754	780	63	1,597	97	56	58

HTML: 754
PDF: 780
XML: 63
Total: 1,597
Supplement: 97
BibTeX: 56
EndNote: 58

Views and downloads (calculated since 25 Feb 2026)

Month	HTML	PDF	XML	Total
Feb 2026	84	60	2	146
Mar 2026	278	258	20	556
Apr 2026	172	177	11	360
May 2026	88	148	3	239
Jun 2026	13	27	3	43
Jul 2026	61	70	4	135
Aug 2026	38	33	8	79
Sep 2026	20	7	12	39

Cumulative views and downloads (calculated since 25 Feb 2026)

Month	HTML	PDF	XML	Total
Feb 2026	84	60	2	146
Mar 2026	278	258	20	556
Apr 2026	172	177	11	360
May 2026	88	148	3	239
Jun 2026	13	27	3	43
Jul 2026	61	70	4	135
Aug 2026	38	33	8	79
Sep 2026	20	7	12	39

Viewed (geographical distribution)

Total article views: 1,593 (including HTML, PDF, and XML) Thereof 1,593 with geography defined and 0 with unknown origin.

Country	#	Views	%

Cited

Latest update: 04 Sep 2026

Download

Preprint (1961 KB)
Metadata XML

Short summary

Persistent clouds and small farm sizes in Southeast Asia have historically hindered accurate mapping. We combined satellite imagery with advanced artificial intelligence to produce the first 10-meter annual cropland map for the region (2019–2024). Achieving over 92 % accuracy, our dataset reveals three to four times more farming on steep slopes than previously reported, providing a vital baseline for monitoring food security and deforestation risks.


Total:	0
HTML:	0
PDF:	0
XML:	0

Annual 10-m high-resolution cropland maps for Southeast Asia since 2019 using AlphaEarth embeddings

Supplement

Data sets

Viewed

Viewed (geographical distribution)

Cited

1 citations as recorded by crossref.