jonghanko/Germany_Wheat_dataset_n_DL_Framework

Germany winter-wheat RSCM dataset & DL framework (2017–2021)

Source-input mirror for: How much ground truth does a satellite-anchored crop yield product need? A controlled sparsification experiment for winter wheat across Germany (GIScience & Remote Sensing).

Citable record = Zenodo, not this repository. The revised paper's reproduction archive — analysis code, district validation panels, the 463 m yield product (2-band cloud-optimized GeoTIFF), all result tables, run manifests, and figures — is deposited on Zenodo, DOI [[TBD: on acceptance]] (COPDESS Open+FAIR). This Hugging Face repository holds the bulky raw inputs and is cross-linked from Zenodo; use Zenodo to cite and to reproduce the paper.

What this study is (and what these files are)

The paper measures how far a process-model (RSCM) yield field anchored to satellite canopy state can substitute for sparse district yield observations — a controlled evaluation, not a prediction study. The files here are the source inputs (satellite/weather/RSCM state variables and the DL framework); the paper's results live on Zenodo.

Honest scope — please read:

• Validation is at the district scale. The pixel-scale yield in Wheat_DEU_dataset.tar.gz uses a reference that is a disaggregation of state statistics, so pixel accuracy is not independently validated; the paper reports accuracy against the 397-district Duden observations (NSE ≈ 0.40).
• The deep-learning emulator's district skill equals static geography (a soil+elevation baseline reproduces it); the district value is in the RSCM product + reliability layer, not the DL surrogate.
• Climate projection is descoped in the paper (its signal is within the emulator error budget).

Contents

File / folder	Role in the revised paper
Wheat_DEU_dataset.tar.gz (~25 GB) — 13 states: weather, MODIS LAI, RSCM state variables, 120-day sequences, pixel yield/GPP/ET	Source inputs. The paper uses the weather/LAI/RSCM-yield streams (aggregated to districts). Pixel yield is used only as the product surface (district-validated).
RSCM_Wheat.tar.gz (~12 GB) — RSCM parameters, observations, simulation outputs	Source inputs (site calibration + simulation).
Scripts_DL_Climate_to_LAI/, Scripts_DL_Climate_n_LAI_to_Yield/, Scripts_DL_RSCM_sim_growth_n_climate_to_Yield/	Original DL framework (LAI retrieval, yield emulation, RSCM-anchored hybrid). The revised analysis code (sparsification, spatial error, uncertainty, district validation, baselines, GATE-3 falsification) is on Zenodo.
Scripts_DL_Climate_to_LAI_CC/ (climate-change scenarios), Scripts_ML_GPP/, Scripts_ML_ET/	NOT results of the revised paper. Projection/scenario and GPP/ET assets belong to a separate, deferred climate-projection study (paper Discussion outlook only). Retained for provenance; do not read them as findings of this paper.
README.md	this file

(If the repository is reorganized, the projection/scenario and GPP/ET items move under a projections_deferred/ prefix; the tarballs stay in place as the raw-input mirror.)

Reproduce the paper

Use the Zenodo archive (code + district panels + product COG + results + manifests + figures) — see its README. Third-party inputs (Duden district yields; MODIS; AgERA5; Thünen crop mask; ISRIC SoilGrids 2.0; Copernicus DEM GLO-90) are referenced there and regenerated by scripts, not re-hosted.

License & citation

Datasets/products: CC BY 4.0. Scripts: MIT. Please cite the paper and the Zenodo DOI ([[TBD]]); this deposit is doi:10.57967/hf/9265. Related: Duden et al. 2024, Sci. Data 11:95 (doi:10.1038/s41597-024-02951-8).