jonghanko/Germany_Wheat_dataset_n_DL_Framework

Germany winter-wheat RSCM source inputs and DL framework

This repository contains historical source-input archives and the original deep-learning scripts for winter wheat in Germany, 2017–2021, together with an aggregate-only September 2026 revision add-on. The related revised manuscript is District yield reference requirements for a satellite-anchored winter wheat product in Germany (submitted to GIScience & Remote Sensing).

Deposits and revision status

The source-input inventory below was checked at repository commit `db9bece7a07de2bf8a38ff01f6fafe0ee99e7543`. This is later than the version recorded in the HF DOI metadata. Cite the relevant deposit and retain the exact commit or archive version used.

The published add-on contains 29 aggregate result files (27 CSV and 2 JSON files) plus 8 documentation, license and manifest files. It covers precipitation and calendar-window comparisons, processed-input quality diagnostics, sensitivity analyses, corrected static-model contrasts, static-model metrics and district-level training-seed spread summaries. The ZIP is 1,093,375 bytes; SHA-256: 207d9a8e9db2d0b80389f619d054a8f95915b3228e3ee1156274bfe2e80bd41c.

The add-on is a results-only supplement, not a complete code-and-data reproduction package. It does not include raw inputs, native crop masks, individual district observations or predictions, model weights, third-party figures, manuscripts or reviewer correspondence. A separately citable code deposit has not been published. The existing HF source files and Zenodo v1.0.0 archive have not been replaced; no new Zenodo version or DOI has been created for this add-on.

Interpretation and validation scope

The revised manuscript evaluates how RSCM pretraining affects district-yield emulation when observed training labels are sparse. It reports gains under specified matched observation budgets, not general replacement of full observation training or universal superiority over static predictors. The additional precipitation and calendar-window comparisons have model-, year- and window-dependent effects; they do not establish a universal precipitation gain or real-time forecasting performance. Those additional revision analyses are not contained in the historical v1.0.0 release.

• —The nominal 463 m spacing describes the output grid. Yield evaluation uses district-level observations; field, parcel and individual-pixel accuracy have not been independently validated. A reference derived by disaggregating state statistics is not independent pixel ground truth.
• —The legacy daily LAI feature is RSCM-simulated LAI. The separately processed VI-derived LAI inputs are distinct from this simulated sequence. Neither finite processed values nor daily simulated slots are counts of native quality-screened satellite acquisitions.
• —The native crop-map identity and the transformation from fine-resolution crop cover to the supplied coarse support remain unresolved. Stored binary masks do not establish fractional wheat cover or pixel purity.
• —The second band of the historical Zenodo yield rasters is district-level A→B emulator training-seed standard deviation mapped to cells, with a state-median fallback. It is not pixel-level error of the yield band, a calibrated prediction interval or full product uncertainty.

Repository contents

The historical analysis code, derived district panels, yield rasters, results and figures are in the Zenodo v1.0.0 ZIP. Its file-to-figure map describes that historical version, not the current manuscript. No individual product TIFF is exposed in the file lists of these two deposits; archive availability is not a demonstration of direct remote COG access.

Licenses and third-party sources

The existing project declarations are CC BY 4.0 for data/products and MIT for scripts. The historical Zenodo ZIP includes LICENSE-CC-BY-4.0 and LICENSE-MIT; retain the applicable notices and attribution when using those materials. The HF repository-level license label is retained from the existing deposit.

Third-party material retains its own source terms. These project-level declarations do not grant rights that the project does not hold. The historical source audit has not established redistribution clearance for every component of the two raw-input archives, including inputs whose native source or permission linkage remains unresolved. This metadata update is not a new rights clearance or a new license for those components.

Relevant source references include Duden et al.'s district-yield data (10.3220/DATA20231117103252-0; data paper), NASA MODIS, Copernicus AgERA5 and DEM, and ISRIC SoilGrids. Consult the original providers' terms and the component-specific provenance before redistribution. The exact native crop-mask source should not be inferred from a citation alone.