Roshangupta007/YieldFormer-DE-data

YieldFormer-DE data

District-level (NUTS-3) data for ten field crops in Germany, 2000 to 2025, used to train and evaluate YieldFormer-DE. Every table can be browsed in the viewer above (choose it in the subset menu) or downloaded.

• Code: https://github.com/roshangupta00750/transformer-based-crop-yield-prediction
• Model and sealed forecasts: https://huggingface.co/Roshangupta007/YieldFormer-DE
• Dashboard: https://yieldformer-de.streamlit.app
• Same files as a download on GitHub: https://github.com/roshangupta00750/transformer-based-crop-yield-prediction/releases/tag/v1.0.0

Authors: Roshan Kumar Gupta, Manisha Samal, Sreedeepthi Ramanath (Julius-Maximilians-Universität Würzburg).

Direct downloads

File	Hugging Face	Same file on GitHub
Training table of the final model	dataset.parquet (624 MB)	dataset_v5_hybridW.parquet
Its policy file	dataset.policy.json	dataset_v5_hybridW.policy.json
Official yields	official_yields.parquet	official_yields.parquet
Backtest forecasts 2010–2025	backtest_forecasts_2010_2025.parquet	backtest_forecasts_2010_2025.parquet
District codes	district_codes.parquet	district_codes.parquet
Checksums		SHA256SUMS.txt

Tables

Table (viewer subset)	File	Rows	Content
official_yields	official_yields.parquet	72,682	Official district yields (Destatis), dt/ha, per district, crop and year, 1999–2025, with district name and state
backtest_forecasts_2010_2025	backtest_forecasts_2010_2025.parquet	37,383	Per district, crop and season: official yield, the YieldFormer-DE backtest forecast, its three members, the XGBoost reference, the 5-year average and the district normal
daily_inputs	dataset.parquet	26.5 million	Training table of the final model: one row per district, crop and day (401 districts, of which 367 have official yields; days from 1 January 2000 to 1 October 2026, yields 2000–2025), with weather, satellite indices, soil, crop development stages and the season's official yield (yield_dt_per_ha)
crop_areas_census	static/crop_area_census.parquet	20,636	Census crop areas per district (hectares)
satellite_arable_land	static/cs_sat.parquet	4.3 million	MODIS NDVI, EVI, LAI and fPAR averaged over arable land, per district and date
crop_development_stages	static/phenology_obs_v3.parquet	1.0 million	Observed crop development stages (DWD), per district, crop and year
terrain	static/dem_features.parquet	401	Elevation, slope and aspect per district
district_codes	district_codes.parquet	475	German district keys (AGS) and NUTS-3 codes, with district name

dataset.parquet is the exact table the final model was trained on (built as dataset_v5_hybridW by scripts/42_build_dataset_v4.py and scripts/55_build_dataset_v5.py --weather hybrid; replaced on 4 November 2026 an earlier build from before the district matching was corrected). Its yields are identical to official_yields and to the official values in backtest_forecasts_2010_2025. To re-run the code, save it with its policy file in data/processed/ as dataset_v5_hybridW.parquet and dataset_v5_hybridW.policy.json, the names the final configs read. Further files in static/ are auxiliary inputs of earlier versions. dataset.policy.json lists the columns of the training table; columns listed as excluded in the code's configs/data.yaml (for example radiation_global) are present but never passed to a model. A backtest forecast is the forecast for a past season from a model trained only on earlier seasons. Yields are in decitonnes per hectare (dt/ha; 1 dt = 100 kg).

Sources and licences

The data combine public sources. Each keeps its own licence; please cite the providers.

Source	Content	Licence
Destatis, Regionalstatistik table 41241-01-03-4	Official district yields	Data Licence Germany, attribution 2.0 (dl-de/by-2-0)
Deutscher Wetterdienst (DWD)	Station weather, crop development stages	CC BY 4.0
Copernicus ERA5-Land	Soil water, radiation	Copernicus licence
NASA MODIS (LP DAAC), via Google Earth Engine	Vegetation indices	NASA open data policy
SoilGrids (ISRIC)	Soil properties	CC BY 4.0
Duden, Nacke and Offermann (2024)	Census crop areas	CC BY 4.0

District boundaries (GISCO, EuroGeographics) are not included; the code downloads them, and their terms do not allow commercial use.