How much ground truth does a satellite-anchored crop yield product need? A controlled sparsification experiment for winter wheat across Germany

Paper: 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: this Zenodo deposit — DOI 10.5281/zenodo.21670326 — openly available under CC BY 4.0 with no embargo and no registration requirement. Source-data mirror: Hugging Face jonghanko/Germany_Wheat_dataset_n_DL_Framework, doi:10.57967/hf/9265 (large raw inputs; see §5). This archive contains the data, code, product, and results needed to reproduce every number, figure, and table in the paper. It is the primary, citable record (COPDESS Open+FAIR); the Hugging Face deposit holds the bulky raw inputs and is cross-linked, not superseded as an input mirror. 1. What this study is (and is not) The study measures, rather than assumes, how far a process-model (RSCM) yield field anchored to satellite canopy state can substitute for sparse district yield observations. It is a controlled evaluation, not a prediction paper. Honest scope — read before use: Validation is at the district scale. The pixel-scale reference is a disaggregation of state statistics, so pixel accuracy is not independently validated; all accuracy statements are district (NSE ≈ 0.40, r ≈ 0.64). Use the per-pixel uncertainty/reliability layer for finer-scale work. Skill is spatial, not interannual (within-year anomaly NSE ≈ 0.03). The DL emulator's district skill = static geography. A soil+elevation baseline reproduces it (GATE-3 test F2b); the district-scale value lives in the RSCM product (+0.084 NSE over that baseline) and the reliability layer, not in the DL surrogate. See the pre-registered falsification gate docs/T18_gate3.md. Climate projection is descoped (its signal lies within the emulator error budget, T11). Projection / scenario assets from the earlier submission are in projections_deferred/ (Hugging Face) and are not a result of this paper. 2. Directory layout code/ src/ tests/ config/ — analysis + reproduction code (MIT) config/paths.example.yaml is the path template; fill it in for your machine data_processed/ district_panel.parquet, district_panel_long.parquet, duden_district_yield.parquet, district_water_features.parquet, trend_model.pkl, agreement_report.csv product/ wheat_yield_DEU_463m_2017..2021.tif — 2-band COG: yield + per-pixel uncertainty (CC BY 4.0) results/ *.csv + manifest/*.json — result tables + per-run reproducibility manifests figures/ Figure1..5.{pdf,png}, FigureS1..S4.{pdf,png} — submission figures docs/ RESULTS_SUMMARY.md, PRODUCT_SPEC.md, T18_gate3.md README.md this file LICENSE-MIT code licence · LICENSE-CC-BY-4.0 data & product licence .zenodo.json deposit metadata 3. File → figure / table map (for reviewers) Figure / Table Source (in this archive) Fig. 1 district validation + support scale results/T05_rscm_district_validation.csv, T13A_scale_support.csv → figures/Figure1.pdf Fig. 2 spatial error + reliability layer results/T13A_spatial_error.csv, T13A_reliability_strata.csv → figures/Figure2.pdf Fig. 3 ground-reference requirement results/T10_sparsification.csv, T10_decision_metrics.csv → figures/Figure3.pdf Fig. 4 anchoring vs observation training results/T09_a_vs_b.csv, T09_residual_inheritance.csv → figures/Figure4.pdf Fig. 5 baseline comparison results/T08_baselines.csv → figures/Figure5.pdf Table 1 validation metrics results/T05_rscm_district_validation.csv Table 2 compute efficiency results/T12A_speedup.csv Suppl. Table S2 static-baseline ladder results/T18_static_baseline_rich.csv Suppl. Fig. S1–S4 figures/FigureS1..S4.pdf (agreement, random sparsification, uncertainty, detrend) GATE-3 falsification (F1–F5, F2b) results/T18_*.csv, results/T18_provenance.csv, docs/T18_gate3.md 4. Reproducibility Each experimental run writes a results/manifest/<run>.json recording git_sha, config_hash, seed, input files, split spec, metrics, wall time, and GPU. Thirty manifests are included; validate them with python -m src.tools.check_manifest ../results/manifest/ (run from code/). Read the following limits before relying on the manifests: Manifests cover runs, not every table. 11 of the 27 CSVs carry no manifest of the same name. Input hashing is incomplete. Of 49 input_files entries, 1 records a SHA-256; the rest name the path only. Use MANIFEST.sha256 at the archive root to verify what you downloaded. Determinism is partial and hardware-bound. 16 manifests record deterministic: true, 13 record false, and 1 does not record the field; determinism is tied to the GPU architecture (NVIDIA sm_80), so exact metric reproduction is expected only on that architecture. Reported metrics are 10-seed means with 95 % confidence intervals. Environment: Python 3.11, uv sync from the pinned pyproject.toml / uv.lock. Regenerate from code/: python -m src.data.rs_product (product COG), then python -m src.viz.make_submission_figures and python -m src.viz.make_single_figures (figures). 5. Third-party inputs — by reference, not re-hosted Input Source In archive Duden 397-district yields Duden et al. 2024, Sci. Data 11:95 (10.1038/s41597-024-02951-8); the data are deposited at OpenAgrar, 10.3220/DATA20231117103252-0, CC BY derived panel only — fetch Final_data.csv manually from the OpenAgrar DOI. No automated downloader ships with this archive: that host fronts downloads with an anti-bot challenge, which a reproduction script should not be circumventing MODIS surface reflectance NASA LP DAAC derived VI/LAI; raw via Hugging Face mirror AgERA5 meteorology Copernicus CDS, CC BY derived district aggregates; regenerable via src/data/agera5_* Thünen crop mask Schwieder et al. 2024 (Zenodo) by reference only — licence-dependent, not re-hosted ISRIC SoilGrids 2.0 Poggio et al. 2021, SOIL 7:217–240 (10.5194/soil-7-217-2021) district means regenerable via src/data/static_features.py Copernicus DEM GLO-90 ESA/Airbus (10.5270/ESA-c5d3d65); attribution: "© DLR e.V. (2014–2018) and © Airbus Defence and Space GmbH" district means regenerable via src/data/static_features.py Large raw inputs (25 GB tar) Hugging Face doi:10.57967/hf/9265 mirror, cross-linked 6. Licence Code (src/, tests/, config/): MIT (LICENSE-MIT). Data, the district panels, and the yield product COG: CC BY 4.0 (LICENSE-CC-BY-4.0). Please cite the paper and this Zenodo DOI.