Yang et al. (2026) Evaluating the Performance of AlphaEarth Foundation Embeddings for Irrigated Cropland Mapping Across Regions and Years

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-04-02
• Authors: Lei Yang, Yan Gao, Xiangyang Zhao, Nannan Liang, Ru Ma, Shixiang Xi, Xiao Zhang, Ruixue Wang
• DOI: 10.3390/rs18071065

Research Groups

Google DeepMind (developer of the AlphaEarth Foundation model)

Short Summary

This study systematically assessed the utility of AlphaEarth Foundation (AEF) model embeddings for irrigated cropland mapping, demonstrating their superior performance in class separability and classification accuracy compared to traditional Sentinel features, with strong temporal transferability.

Objective

• To comprehensively assess AlphaEarth Foundation (AEF) embeddings for irrigated cropland mapping in terms of feature separability, classification performance, and spatiotemporal transferability.

Study Configuration

• Spatial Scale: Regional scale, specifically the Guanzhong Plain in China and Kansas in the USA.
• Temporal Scale: Data from 2022 and 2024 for cross-year transfer experiments.

Methodology and Data

• Models used: K-means clustering, Decision Tree (DT), Random Forest (RF), Gradient Boosting Decision Trees (GBDT), Support Vector Machine (SVM), Multi-layer Perceptron (MLP).
• Data sources: AlphaEarth Foundation (AEF) model embeddings (64 dimensions), Sentinel-2 bands, Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Normalized Difference Water Index (NDWI), Sentinel-1 vertical transmit vertical receive (VV) backscatter, and Sentinel-1 vertical transmit horizontal receive (VH) backscatter.

Main Results

• AEF embeddings consistently provided stronger class separability in both study areas, with a maximum Jeffries–Matusita (JM) distance of 1.58 (dimension A29).
• Using AEF embeddings, the Random Forest (RF) classifier achieved overall accuracies (OA) of 0.95 in the Guanzhong Plain and 0.93 in Kansas, outperforming models based on Sentinel-1/2 bands and indices.
• Unsupervised K-means clustering on AEF embeddings yielded overall accuracies greater than 0.85, indicating high intrinsic separability between irrigated and rainfed croplands.
• Transfer experiments demonstrated stable temporal transferability (cross-year OA > 0.87).
• Cross-region transferability was limited (OA approximately 0.3) due to differences in irrigation regimes, crop phenology, and management practices.

Contributions

• First systematic assessment of AlphaEarth Foundation (AEF) embeddings for irrigated cropland mapping.
• Demonstrated the significant potential of high-information-density representations from geospatial foundation models for this application.
• Provided methodological and technical insights to support transfer learning and operational mapping of irrigated croplands over large areas.

Funding

• Not specified in the provided text.

Citation

@article{Yang2026Evaluating,
  author = {Yang, Lei and Gao, Yan and Zhao, Xiangyang and Liang, Nannan and Ma, Ru and Xi, Shixiang and Zhang, Xiao and Wang, Ruixue},
  title = {Evaluating the Performance of AlphaEarth Foundation Embeddings for Irrigated Cropland Mapping Across Regions and Years},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18071065},
  url = {https://doi.org/10.3390/rs18071065}
}