Li et al. (2026) An accurate 10 m annual crop map product of maize and soybean across the United States

Identification

• Journal: Earth system science data
• Year: 2026
• Date: 2026-03-25
• Authors: Haijun Li, Xiao‐Peng Song, Bernard Adusei, Jeffrey Pickering, Andre Lima, Andrew Poulson, Antoine Baggett, Peter Potapov, Ahmad Khan, Viviana Zalles, Andres Hernandez-Serna, Samuel M. Jantz, Amy Pickens, Carolina Ortiz-Dominguez, Xinyuan Li, Theodore Kerr, Zhen Song, Svetlana Turubanova, Eddy Bongwele, Héritier Koy Kondjo, Anna Komarova, Stephen V. Stehman, Matthew C. Hansen
• DOI: 10.5194/essd-18-2227-2026

Research Groups

• Department of Geographical Sciences, University of Maryland, College Park, MD, United States
• World Resources Institute, Washington, DC, United States
• National Institute for Modeling Biological Systems, University of Tennessee, Knoxville, TN, United States
• Department of Sustainable Resources Management, SUNY College of Environmental Science and Forestry, Syracuse, NY, United States

Short Summary

This study developed an openly available, annual, 10 m spatial resolution maize and soybean map product for the Contiguous United States (CONUS) from 2019 to 2022, achieving consistent overall accuracies greater than 95%. The research demonstrates that these higher-resolution maps significantly reduce mixed pixels compared to existing 30 m products, enhancing agricultural monitoring capabilities.

Objective

• To develop annual 10 m spatial resolution crop maps of maize and soybean across the Contiguous United States (CONUS) using Sentinel-2 time series data.
• To quantify the benefits of these 10 m maps in reducing mixed pixels compared to existing 30 m products.

Study Configuration

• Spatial Scale: Contiguous United States (CONUS), with a pixel resolution of 10 meters.
• Temporal Scale: Annual maps from 2019 to 2022, using Sentinel-2 data acquired between May and October for each year.

Methodology and Data

• Models used:
  • Decision Tree classifiers (for PSU-level crop mapping)
  • Random Forest (RF) classifiers (for national-scale wall-to-wall crop classification)
• Data sources:
  • Satellite: Sentinel-2A and -2B Level-2A Bottom of the Atmosphere reflectance (S2 L2A) data (10 m and 20 m bands) from Google Cloud.
  • Observation: Annual field surveys and ground data collected using a stratified, two-stage cluster sampling design (probability sample for validation, "windshield survey" for training).
  • Ancillary Data:
    • TanDEM-X data (nominal 12 m, resampled to 10 m) for elevation, slope, and aspect.
    • US Department of Agriculture (USDA) National Agricultural Statistics Service (NASS) Cropland Data Layer (CDL) for comparison.
    • USDA NASS agricultural statistics (county and state levels) for comparison.
    • US TIGER database for road and rail networks.

Main Results

• The annual 10 m maize and soybean maps for CONUS (2019-2022) achieved consistent overall accuracies greater than 95%, with standard errors less than 1%.
• User's accuracies (UAs) and producer's accuracies (PAs) for maize were consistently higher than 91% and 84%, respectively.
• UAs and PAs for soybean were consistently greater than 88% and 82%, respectively.
• Comparison with USDA NASS official statistics showed close agreement: state-level r² > 0.99 (RMSD < 900 km² for maize, < 1800 km² for soybean); county-level r² > 0.97 (RMSD between 30 and 50 km²).
• The 10 m maps significantly reduced mixed pixels compared to 30 m resolution. Aggregating the 10 m maps to 30 m resolution showed a median reduction of 8% for maize and 9% for soybean mixed pixels across all counties.
• The 10 m maps provided clearer field boundaries and captured more landscape fragmentation, especially in regions with smaller, more diverse fields or complex agricultural/wetland mosaics.

Contributions

• Developed the first openly available, annual, 10 m spatial resolution maize and soybean map product for the entire Contiguous United States (CONUS) from 2019 to 2022.
• Demonstrated a robust workflow for large-area, high-resolution crop mapping that integrates Sentinel-2 Analysis Ready Data (ARD) generation, stratified two-stage cluster sampling for field data collection, and machine learning (Random Forest).
• Quantified the substantial benefits of 10 m resolution mapping over 30 m products in reducing mixed pixels, particularly outside of the major Corn Belt regions.
• Showcased the robustness of using multi-temporal metrics to normalize crop phenological variations, enabling consistent annual map production despite inter- and intra-annual phenological shifts.
• The generated maps are available earlier in the growing season (3-4 months) than official 30 m products, providing timely information for agricultural monitoring.

Funding

• Google Academic Research Awards
• NASA Land-Cover and Land-Use Change Program (80NSSC24K0188, 80NSSC23K0526)
• NASA Acres Program (80NSSC23M0034)
• Bezos Earth Fund
• University of Maryland

Citation

@article{Li2026accurate,
  author = {Li, Haijun and Song, Xiao‐Peng and Adusei, Bernard and Pickering, Jeffrey and Lima, Andre and Poulson, Andrew and Baggett, Antoine and Potapov, Peter and Khan, Ahmad and Zalles, Viviana and Hernandez-Serna, Andres and Jantz, Samuel M. and Pickens, Amy and Ortiz-Dominguez, Carolina and Li, Xinyuan and Kerr, Theodore and Song, Zhen and Turubanova, Svetlana and Bongwele, Eddy and Kondjo, Héritier Koy and Komarova, Anna and Stehman, Stephen V. and Hansen, Matthew C.},
  title = {An accurate 10 m annual crop map product of maize and soybean across the United States},
  journal = {Earth system science data},
  year = {2026},
  doi = {10.5194/essd-18-2227-2026},
  url = {https://doi.org/10.5194/essd-18-2227-2026}
}