jonghanko/Cornbelt_dataset_n_Deep_Learning_Framework

US Corn Belt Maize Yield Dataset and Deep Learning Framework (2012–2020)

This repository contains the dataset and deep-learning scripts associated with the study:

Jeong, S., Ko, J., Shin, T., Ban, J.-O., Wie, J., Yeom, J.-M. Integrating deep learning and satellite imagery for spatiotemporal maize yield prediction in the US Corn Belt. International Journal of Applied Earth Observation and Geoinformation (submitted).

Overview

We pair optimization-based assimilation of MODIS-derived leaf area index (LAI) into the process-based Remote Sensing-integrated Crop Model (RSCM) with five deep-learning regressors — Feed-Forward Neural Network (FFNN), Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Gated Recurrent Unit (GRU), and Transformer — to predict state-level maize yield at 500-m spatial resolution across seven US Corn Belt states (Iowa, Illinois, Indiana, Minnesota, Nebraska, Ohio, South Dakota) for 2012–2020. The GRU produced the most stable out-of-sample performance (mean NSE = 0.92 on the 2020 holdout).

The repository provides:

1. Core dataset (~271 GB, zipped): processed MODIS reflectance and land-surface temperature, AgERA5 meteorology, and CORDEX-North America climate projections at 500-m resolution across the seven-state domain.
2. Analysis scripts (unzipped): training, inference, and visualization code for the three modeling workflows described in the paper.

Repository structure

• Cornbelt_dataset.zip — Processed inputs (folder name inside archive: MODIS_CornBelt_2012_to_2020):
  • MODIS MOD09A1 Collection 6.1 (8-day surface reflectance, 500 m)
  • MODIS MOD11A1 Collection 6.1 (daily land-surface temperature, resampled from 1 km to 500 m)
  • AgERA5 daily meteorology (downward surface solar radiation, 2-m maximum and minimum air temperature), resampled from 0.1° (~9 km) to 500 m
  • CORDEX-North America projections from GERICS-REMO2015 forced by MPI-ESM-LR and NorESM1-M, under RCP 2.6 and RCP 8.5, for the baseline (2006–2025), 2050s (2040–2060), and 2090s (2080–2100)
• scripts_Climate_n_LAI/ — LAI estimation from climate drivers (daily max/min temperature and solar radiation)
• Scripts_RSCM_sim_growth_n_climate_to_Yield/ — yield prediction and spatial mapping using RSCM-simulated growth variables combined with climate inputs (hybrid RSCM-ML configuration; used for yield validation in the paper)
• Scripts_Climate_n_LAI_to_Yield/ — yield prediction and spatial mapping from climate inputs only (configuration used for future CORDEX-driven projections)

Dataset details

• Spatial coverage: seven US Corn Belt states — Iowa, Illinois, Indiana, Minnesota, Nebraska, Ohio, South Dakota (~1.5 million km², ~1.1 million 500-m cropland pixels)
• Spatial resolution: 500 m, Albers Equal Area Conic projection
• Temporal range: 2012–2020 (historical); 2006–2025, 2040–2060, and 2080–2100 (CORDEX projections)
• Key variables: Leaf Area Index (LAI), above-ground biomass, maize yield, NDVI, MTVI1, OSAVI, RDVI, solar radiation, maximum and minimum air temperature, land-surface temperature
• Reference yields: USDA National Agricultural Statistics Service (USDA-NASS) state-level maize yield, 2012–2020 (63 state-year combinations)

Intended use

This dataset is suitable for research on regional-scale crop-yield prediction, remote-sensing–based agroecosystem monitoring, hybrid process-based + machine-learning modeling, and climate-change impact assessment for maize systems. It is intended for research and educational purposes.

Out-of-scope use

The 500-m "observed yield" maps were produced by disaggregating USDA-NASS state totals in proportion to RSCM-simulated pixel yield and are not independent pixel-level observations. Quantitative accuracy statements in the paper are therefore made at the state-year aggregation level. Users should not treat the pixel-level reference maps as ground truth; county-level USDA-NASS survey data (not included here) provide statistically independent validation.

The climate-projection outputs use a climate-only input configuration that differs from the validated hybrid RSCM-ML configuration; projections should be read as indicative scenario analyses rather than calibrated forecasts. CO₂ fertilization and irrigation management are not represented.

Installation and usage

Download via huggingface-cli:

huggingface-cli download <HF-USERNAME>/<HF-REPO-NAME> --repo-type dataset

Unzip the core dataset archive and run the scripts in the relevant workflow directory. See individual README files inside each Scripts_* folder for environment setup and execution instructions. The scripts were developed against Python 3.8 and PyTorch 1.13.1 and require a CUDA-capable GPU for training (NVIDIA A100 used in the paper).

License

Dataset (Cornbelt_dataset.zip): CC BY 4.0. Scripts (scripts_*/ directories): MIT License (see LICENSE file in each script directory).

Citation

If you use this dataset or the accompanying scripts, please cite:

@article{Jeong2026CornBelt,
  author  = {Jeong, Seungtaek and Ko, Jonghan and Shin, Taewhan and Ban, Jong-oh and Wie, Jieun and Yeom, Jong-Min},
  title   = {Integrating deep learning and satellite imagery for spatiotemporal maize yield prediction in the US Corn Belt},
  journal = {Nature Food},
  year    = {2026},
  note    = {Submitted}
}

Please also cite the underlying data providers:

• MODIS MOD09A1 / MOD11A1: NASA LP DAAC
• AgERA5: Boogaard et al. (2020), ECMWF Copernicus Climate Change Service
• CORDEX-North America: Copernicus Climate Change Service (2020), doi:10.24381/cds.bc91edc3
• USDA-NASS Quick Stats: https://quickstats.nass.usda.gov/

Contact

Jonghan Ko (corresponding author) Applied Plant Science, Chonnam National University, Gwangju, South Korea Email: jonghan.ko@jnu.ac.kr

Acknowledgements

See the acknowledgements section of the associated manuscript.