Liu et al. (2026) From national to global: Mapping global center-pivot irrigated cropland with knowledge-guided data augmentation

Identification

• Journal: Remote Sensing of Environment
• Year: 2026
• Date: 2026-10-12
• Authors: Xiangyu Liu, Wei He, Yu Xia, Jie Wang, Hongyan Zhang
• DOI: 10.1016/j.rse.2026.115736

Research Groups

• Pengcheng Laboratory, Shenzhen, China
• State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, Hubei, China
• School of Computer Science, China University of Geosciences, Wuhan, China

Short Summary

This study developed a novel deep learning framework integrating prior knowledge to generate a global training dataset for mapping center-pivot irrigated cropland (CPIC) in 2020, revealing a significantly underestimated global CPIC area and its distribution in water-stressed regions.

Objective

• To develop a novel methodological framework integrating CPIC prior knowledge and deep learning to produce a 2020 global CPIC map (GCMap) based on Landsat imagery.
• To assess the global distribution and area of CPIC, particularly in relation to groundwater recharge rates.

Study Configuration

• Spatial Scale: Global
• Temporal Scale: 2020 (for the map generation), utilizing multi-temporal imagery.

Methodology and Data

• Models used: Deep learning (teacher model, student model), noise-resistant loss function, shape-prior guided Projection Simulation (PS).
• Data sources: Landsat imagery, Google Earth Engine cloud computing platform, national-scale CPIC samples, global CPIC pseudo-labels.

Main Results

• The global error-adjusted area of CPIC in 2020 was estimated at 18.70 ± 0.42 Mha, representing a 63% increase relative to previous estimates (11.5 Mha).
• The GCMap achieved a producer's accuracy (PA) of 0.86 ± 0.02, indicating a marked reduction in omission errors compared to existing CPIC products.
• Geographically, the semi-arid zone accounts for the largest share of CPIC (9.27 ± 0.29 Mha, 49.28%), followed by the arid zone (4.69 ± 0.20 Mha, 24.95%) and humid zone (2.99 ± 0.22 Mha, 15.92%).
• Approximately 15.25 ± 0.87 Mha (82.04%) of CPIC areas are located in regions with groundwater recharge rates below 100 mm/year, with 9.97 ± 0.48 Mha (53.62% of total CPIC) falling within extremely low recharge zones (<20 mm/year).

Contributions

• Developed a novel methodological framework that integrates CPIC prior knowledge into the generation of a global training dataset for CPIC mapping.
• Proposed a shape-prior guided Projection Simulation (PS) to enhance the diversity of national-scale CPIC samples for global pseudo-label generation.
• Implemented a stratified sampling strategy and a dual-temporal correction mechanism to generate high-confidence global training samples.
• Produced a high-resolution (Landsat-based) 2020 global CPIC map (GCMap) with statistically rigorous area estimates and significantly reduced omission errors.
• Revealed a substantial underestimation of global CPIC area in previous estimates and highlighted its predominant distribution in aquifers with limited groundwater recharge.
• The GCMap inventory is publicly available for further research and policymaking.

Funding

[Information not available in the provided paper text.]

Citation

@article{Liu2026From,
  author = {Liu, Xiangyu and He, Wei and Xia, Yu and Wang, Jie and Zhang, Hongyan},
  title = {From national to global: Mapping global center-pivot irrigated cropland with knowledge-guided data augmentation},
  journal = {Remote Sensing of Environment},
  year = {2026},
  doi = {10.1016/j.rse.2026.115736},
  url = {https://doi.org/10.1016/j.rse.2026.115736}
}