County-Level US Corn Yield Lead-Time Forecasting Benchmark (Corn Belt, 2000–2025)

This dataset is a county-year modeling matrix for benchmarking in-season corn yield forecasts across the US Corn Belt, assembled entirely from public data products. It supports the study of how early a county-level corn yield forecast can be trusted, and enables reproducible feature-ablation and lead-time forecasting experiments. The matrix contains 22,797 county-year records (632 columns) spanning 2000–2025 across 12 Corn Belt states (IL, IN, IA, KS, MI, MN, MO, NE, ND, OH, SD, WI), with counties identified by 5-digit FIPS codes. The prediction target is raw USDA NASS county corn grain yield (yield_bu_acre, bushels/acre). Features are organized into groups derived from public sources: monthly and weekly weather (GRIDMET); heat stress (GDD/KDD/EDD); drought and water balance (GRIDMET drought, TerraClimate); satellite vegetation and water indices (MODIS NDVI/EVI/NDWI/GCI/LAI/FPAR/LST); multi-depth soil (POLARIS, OpenLandMap); irrigation fraction (LGRIP30); geography (US Census, SRTM); USDA NASS crop progress/condition reports; and lagged historical yield. Full per-column documentation is provided in the accompanying data dictionary, and upstream product/asset IDs, native resolution, and aggregation methods are listed in the source-provenance file. The associated publication analyses the 2000–2024 subset with a valid reported yield (22,138 county-years across 1,009 reporting units); the released loader reproduces this scope. The complete matrix (2000–2025) is provided here as a faithful pipeline artifact so users can apply their own filtering. Files: benchmark_matrix.parquet (primary), benchmark_matrix.csv.gz (fallback), benchmark_matrix_dictionary.csv, benchmark_matrix_build_report.csv, source_provenance.csv, plus README_DATA.md, LICENSE, and MANIFEST.txt (SHA-256 checksums). Note: yield_trend and yield_anomaly are full-record descriptive columns and must not be used as forecast inputs; use the past-only trend feature provided by the reproduction code. Released under CC BY 4.0 — please cite both the dataset and the associated paper.