Jiang et al. (2025) CropLayer: a 2 m resolution cropland map of China for 2020 from Mapbox and Google satellite imagery

Identification

• Journal: Earth system science data
• Year: 2025
• Date: 2025-12-03
• Authors: Hao Jiang, Mengjun Ku, Xia Zhou, Zheng Qiong, Yangxiaoyue Liu, Jianhui Xu, Dan Li, Chongyang Wang, Jiayi Wei, Jing Zhang, Shuisen Chen, Jianxi Huang
• DOI: 10.5194/essd-17-6703-2025

Research Groups

• Guangdong Engineering Technology Research Center of Remote Sensing Big Data Application, Key Lab of Guangdong for Utilization of Remote Sensing and Geographical Information System, Guangzhou Institute of Geography, Guangdong Academy of Sciences, Guangzhou, China
• Department of Geomatics Engineering, School of Traffic & Transportation Engineering, Changsha University of Science & Technology, Changsha, China
• State Key Laboratory of Resources and Environmental Information Systems, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, China
• Faculty of Geosciences and Engineering, Southwest Jiaotong University, Chengdu, China
• College of Land Science and Technology, China Agricultural University, Beijing, China

Short Summary

This study presents CropLayer, a 2 meter resolution cropland map of China for 2020, developed from Mapbox and Google satellite imagery. It achieves high accuracy (pixel-level 88.73%, block-level 96.5%) and strong consistency with official statistics, with 30 out of 32 provincial units showing area estimates within ±10% deviation, significantly outperforming existing datasets.

Objective

• To develop a high-resolution (2 meter) cropland map of China for 2020 (CropLayer) by integrating Mapbox and Google satellite imagery.
• To address limitations of existing cropland products for China, such as low consistency, inaccurate boundary delineation, and extreme area estimation deviations, particularly for smallholder farming systems.

Study Configuration

• Spatial Scale: Entirety of China (approximately 9.6 million square kilometers). The final map resolution is 2 meters. Imagery was processed in 0.05° × 0.05° blocks.
• Temporal Scale: Cropland map for the year 2020. High-resolution imagery was accessed between August 2022 and December 2023.

Methodology and Data

• Models used:
  • ResNet (for Image Quality Assessment - IQA)
  • Mask2Former (for semantic segmentation in cropland extraction)
  • XGBoost (for semantic correctness evaluation and results integration)
• Data sources:
  • High-resolution satellite imagery: Mapbox and Google satellite imagery (approximately 2.4 meter resolution, level-16, RGB color).
  • DEM data: 30 meter resolution Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (for deriving slope, ruggedness, roughness).
  • Samples: Manually labeled cropland samples (157,395 polygons across 366 image blocks), negative cropland samples (761 blocks), image blocks for coverage type classification (8084 blocks), and semantic correctness/results integration (3891 image blocks).
  • Validation data: Provincial-level cropland statistics from the Third National Land Survey (TNLS) for China.
  • Comparison datasets: Eight existing cropland/land-cover datasets (CACD, CLCD, WorldCover, ESRI Land Cover v2, FROM, FCS30, Globeland30-2020, SinoLC).

Main Results

• The ResNet-based Image Quality Assessment (IQA) model achieved a top-1 accuracy of 95.6% for classifying image coverage types. Mapbox and Google imagery showed complementary spatial distributions of low-quality regions.
• The Mask2Former model achieved the highest pixel-level Intersection over Union (IoU) of 88.73% for cropland segmentation.
• The XGBoost classifier for block-level semantic correctness achieved an overall accuracy of 94.3%, with combined true positive and true negative predictions reaching 96.5% (95.9% for Mapbox, 97.1% for Google).
• CropLayer demonstrated strong consistency with official provincial statistics, with 30 out of 32 provinces showing cropland area estimates within ±10% deviation. In contrast, existing datasets met this criterion for only 1 to 9 provinces.
• Pixel-level IoU comparisons with existing datasets ranged from 0.44 to 0.62, with ESA showing the highest agreement (0.62).
• Block-level analysis revealed that deviations in area fraction and edge density between CropLayer and coarser products were concentrated in mountainous and hilly regions (e.g., Yunnan-Guizhou Plateau), where coarser maps tend to simplify boundaries and merge small plots. Edge-density underestimation was most pronounced in the 10-25° slope range.
• Feature importance analysis for the integration model highlighted Area Fraction (Google), Area (Google), Slope, and Area (Mapbox) as the most influential features, indicating the critical roles of spatial extent and topographic context in data fusion.

Contributions

• Developed CropLayer, a novel 2 meter resolution cropland map of China for 2020, which significantly improves accuracy and consistency with official statistics compared to existing datasets.
• Introduced a robust three-stage workflow incorporating ResNet-based Image Quality Assessment, an active learning strategy with Mask2Former for semantic segmentation, and XGBoost-based integration of Mapbox and Google imagery.
• Implemented a comprehensive three-level validation scheme (pixel, block, and regional scales) that served as objective stopping criteria for active learning, ensuring robust, interpretable accuracy and mitigating overfitting.
• Demonstrated the effectiveness of combining complementary high-resolution satellite imagery sources for large-scale, detailed cropland mapping, particularly in fragmented and topographically complex terrains.
• Provided a transparent and reproducible benchmark dataset for agricultural monitoring, yield estimation, and land-use studies in China, supporting advancements in agricultural modeling and AI-based analytics.

Funding

• National Natural Science Foundation of China (grant no. 42071417)
• GDAS' Project of Science and Technology Development (grant no. 2022GDASZH-2022010102)
• Guangzhou Science and Technology Plan Project (grant no. 2023B03J1373)
• Hunan Provincial Natural Science Foundation of China (grant no. 2023JJ40025)

Citation

@article{Jiang2025CropLayer,
  author = {Jiang, Hao and Ku, Mengjun and Zhou, Xia and Qiong, Zheng and Liu, Yangxiaoyue and Xu, Jianhui and Li, Dan and Wang, Chongyang and Wei, Jiayi and Zhang, Jing and Chen, Shuisen and Huang, Jianxi},
  title = {CropLayer: a 2 m resolution cropland map of China for 2020 from Mapbox and Google satellite imagery},
  journal = {Earth system science data},
  year = {2025},
  doi = {10.5194/essd-17-6703-2025},
  url = {https://doi.org/10.5194/essd-17-6703-2025}
}