jonghanko/Germany_Barley_dataset_n_DL_Framework

Germany Barley (Yield, GPP, & ET) Dataset and Deep Learning Framework (2017–2021)

This repository contains the processed dataset and analysis scripts for winter-barley modeling across 13 German states (2017–2021), including yield, gross primary production (GPP), evapotranspiration (ET), and climate-change scenario outputs.

Revision notes (2026-07). A post-submission audit (see Revision_CEEA_2026/) established that the modeled crop is winter barley, not spring barley — the nationwide pixels match the Thünen/Schwieder winter.barley class mask exactly and the RSCM was run with winter-barley parameters. Earlier "spring-barley" wording in this README has been corrected. The audit also (i) reproduced the published metrics and flagged corrections (ET NSE, a GPP unit label), (ii) added an independent 397-district validation (Düden et al. 2024) showing skill is strongly scale-dependent (state NSE ≈0.9 vs district ≈0.4), (iii) quantified the two-stage climate→LAI→yield emulator's error and found the national-mean climate-change signal is within that error, and (iv) added an independent VPD water-stress diagnostic. See Revision_CEEA_2026/ for the full documentation, results tables, figures, and analysis code. The 500-m "reference" yields are state statistics disaggregated by RSCM pixel yield — treat quantitative claims at the state/district aggregation level (see Out-of-scope use).

Overview

We combine MODIS-based remote sensing, AgERA5 meteorology, and CORDEX-Europe climate projections with the process-based Remote Sensing-integrated Crop Model (RSCM) and deep-learning / machine-learning regressors to simulate spatiotemporal barley yield, GPP, and ET at 500-m resolution.

The repository provides:

1. Processed dataset archives (~26 GB compressed):
   • Barley_DEU_dataset.tar.gz (~17 GB) — state-level MODIS, weather, LAI, RSCM growth, yield, GPP, ET, and climate-change outputs for 13 German states.
   • RSCM_Barley.tar.gz (~9 GB) — RSCM barley model parameter files, observation inputs, and simulation outputs used to build the hybrid modeling workflow.
2. Analysis scripts (unzipped): training, inference, and visualization code for LAI estimation, yield prediction, GPP simulation, ET simulation, and climate-change analyses.
3. Revision_CEEA_2026/ (new): the post-submission audit and additional analyses — documentation, results tables, figures, and code (no large data; derived products only).

Repository structure

<repo-root>/
├── README.md
├── Barley_DEU_dataset.tar.gz
├── RSCM_Barley.tar.gz
├── Scripts_DL_Climate_to_LAI/
├── Scripts_DL_Climate_to_LAI_CC/
├── Scripts_DL_Climate_n_LAI_to_Yield/
├── Scripts_DL_RSCM_sim_growth_n_climate_to_Yield/
├── Scripts_ML_GPP/
├── Scripts_ML_ET/
└── Revision_CEEA_2026/          # audit docs, results, figures, analysis code

Barley_DEU_dataset.tar.gz

Top-level folder inside the archive: Barley_DEU_dataset/, with one subdirectory per state:

BadenW, Bayern, Brandenburg, Hessen, MecklenburgV, Niedersachsen, NordrheinW, RheinlandP, Saarland, Sachsen, SachsenA, SchleswigH, Thuringen

Typical contents per state include:

• weather/ — AgERA5-based daily meteorology resampled to the 500-m grid (solar radiation, Tmax, Tmin)
• LAI_VIs/ — MODIS-derived vegetation indices and LAI-related inputs
• processed_RSCM_data/ — RSCM growth variables prepared for ML workflows
• data_LAI_geo_wx_2017_to_21/ — 120-day LAI + weather + geolocation .npy sequences (historical 2017–2021)
• data_LAI_geo_wx_CC2050_RCP26/, ..._CC2050_RCP85/, ..._CC2070_RCP26/, ..._CC2070_RCP85/, ..._CC2090_RCP26/, ..._CC2090_RCP85/ — climate-change perturbed LAI + weather inputs (the LAI channel is the FFNN climate→LAI projection; see Revision_CEEA_2026/)
• coord_n_DNN_sim_yield/ — reference and simulated yield arrays at the pixel level
• out_pixGro1_yr2017/ … out_pixGro1_yr2021/ — pixel-level RSCM growth outputs by year
• data_GPP_2017_to_21/, data_ET_2017_to_21/ — historical GPP and ET simulation outputs
• data_GPP_CC2050_RCP26_ML/, data_ET_CC2050_RCP26_ML/, … — ML-simulated GPP and ET under climate-change scenarios
• {State}_map/ — state boundary shapefiles for mapping
• vis/ — visualization outputs

Historical LAI/weather .npy files use shape (P, 120, 8) with channels:

[DOY1, LAI, Easting, Northing, DOY2, solar radiation, Tmax, Tmin]

RSCM_Barley.tar.gz

Top-level folder inside the archive: RSCM_Barley/, containing RSCM winter-barley model setup files, parameter tables, LAI observation files, and state/year simulation outputs used in the assimilation and hybrid yield workflows. The nationwide application directory RSCM_Barley/2017_to_2021_MODIS_Germany_Barley/{State}/class_map/ holds the per-year Thünen/Schwieder crop masks (DEU.winter.barley.*.bin, DEU.spring.barley.*.bin, DEU.winter.wheat.*.bin); the modeled pixels correspond to DEU.winter.barley. Calibration draws on winter-barley (Sites 5, 6) and, in smaller part, spring-barley (Site 4) eddy-covariance/field data from Baden-Württemberg (Kraichgau, Swabian Alb).

Script folders

• Scripts_DL_Climate_to_LAI/ — deep-learning LAI estimation from daily weather drivers (FFNN, LSTM, BiLSTM, GRU, Transformer) for each German state
• Scripts_DL_Climate_to_LAI_CC/ — climate-change LAI projection and seasonal / delta visualization
• Scripts_DL_Climate_n_LAI_to_Yield/ — yield prediction from climate + LAI inputs (walk-forward CV, CC summary plots)
• Scripts_DL_RSCM_sim_growth_n_climate_to_Yield/ — hybrid RSCM growth + climate yield modeling (temporal and spatiotemporal configurations)
• Scripts_ML_GPP/ — machine-learning GPP simulation from RSCM growth and weather variables, including CC application scripts
• Scripts_ML_ET/ — machine-learning ET simulation from RSCM growth and weather variables, including CC application scripts

Dataset details

• Crop: winter barley (Germany)
• Spatial coverage: 13 German states — Baden-Württemberg, Bayern, Brandenburg, Hessen, Mecklenburg-Vorpommern, Niedersachsen, Nordrhein-Westfalen, Rheinland-Pfalz, Saarland, Sachsen, Sachsen-Anhalt, Schleswig-Holstein, and Thüringen
• Spatial resolution: 500 m
• Temporal range: 2017–2021 (historical); climate-change scenarios for the 2050s, 2070s, and 2090s under RCP 2.6 and RCP 8.5
• Key variables: leaf area index (LAI), above-ground biomass / growth, barley yield, GPP, ET, solar radiation, maximum and minimum air temperature, vegetation indices
• Reference yields: state-level official statistics disaggregated to the 500-m cropland grid following the RSCM-based downscaling workflow used in the associated study

Intended use

This dataset is suitable for research on regional-scale winter-barley yield prediction, GPP and ET modeling, remote-sensing-based agroecosystem monitoring, hybrid process-based + machine-learning modeling, and climate-change impact assessment for barley systems in Germany. It is intended for research and educational purposes.

Out-of-scope use

The 500-m reference yield maps were produced by disaggregating state-level totals in proportion to RSCM-simulated pixel yield and are not independent pixel-level observations. Quantitative accuracy statements should therefore be interpreted at the state-year aggregation level unless independently validated (see the district-scale validation in Revision_CEEA_2026/, which shows skill drops markedly from state to district scale).

Climate-change outputs based on climate-only or ML-transfer configurations should be read as scenario analyses rather than calibrated forecasts. The projection is a two-stage climate→LAI→yield emulator whose compounded error is comparable to the national-mean signal (see Revision_CEEA_2026/); CO₂ fertilization effects and irrigation management are not represented.

Installation and usage

Download the repository with the Hugging Face CLI:

huggingface-cli download jonghanko/Germany_Barley_dataset_n_DL_Framework --repo-type dataset --local-dir ./Germany_Barley_dataset

Extract the archives:

cd Germany_Barley_dataset
tar -xzf Barley_DEU_dataset.tar.gz
tar -xzf RSCM_Barley.tar.gz

Run the scripts in the relevant workflow directory. Deep-learning training scripts were developed with Python 3.10+ and PyTorch and require a CUDA-capable GPU for model training. Machine-learning GPP/ET scripts use scikit-learn, XGBoost, and LightGBM.

Example:

cd Scripts_DL_Climate_n_LAI_to_Yield
python main.py

License

Dataset archives (Barley_DEU_dataset.tar.gz, RSCM_Barley.tar.gz): CC BY 4.0.

Scripts (Scripts_*/ directories) and Revision_CEEA_2026/ code: MIT License (see LICENSE in each script directory, where provided).

Third-party data used only as external references in Revision_CEEA_2026/ (not redistributed here): the German 397-district yield dataset (Düden, Nacke & Offermann 2024, Sci Data 11:95, OpenAgrar doi:10.3220/DATA20231117103252-0) and AgERA5 vapour pressure (Copernicus CDS). Please obtain these from their original providers under their own licenses.

Citation

If you use this dataset or the accompanying scripts, please cite the associated manuscript and the underlying data providers below.

@article{Ko2026GermanyBarley,
  author  = {Ko, Jonghan and Shawon, Ashifur Rahman and Jeong, Seungtaek and Shin, Taewhan},
  title   = {Germany winter barley dataset and deep-learning framework (2017--2021)},
  year    = {2026},
  note    = {Manuscript under revision}
}

Please also cite the underlying data providers:

• MODIS MOD09A1 / MOD11A1: NASA LP DAAC
• AgERA5: Boogaard et al. (2020), ECMWF Copernicus Climate Change Service
• CORDEX-Europe: Copernicus Climate Change Service, doi:10.24381/cds.bc91edc3
• German 397-district crop yields (revision validation): Düden, Nacke & Offermann (2024), Scientific Data 11:95, doi:10.1038/s41597-024-02951-8

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.