2026-10-14T18:53:10Z https://www.scidb.cn/oai

10.57760/sciencedb.0133e 2026-09-24T09:46:06Z

2026-09-24 training_samples_zarr_v34.zip doi:10.57760/sciencedb.0133e en This dataset is a multi-scale spatiotemporal foundation benchmark designed for field-level and county-level crop yield prediction across the contiguous United States (CONUS). The dataset integrates multi-source Earth observation records, combining 16-day MODIS MOD13C1 vegetation indices (NDVI), ERA5-Land surface meteorological reanalysis (temperature, precipitation, downward solar radiation, GDD, EDD, VPD, extreme climate indices, and dual-layer volumetric soil moisture SWVL1/SWVL2), USDA Cropland Data Layer (CDL) crop-specific geographic masks, and official USDA county-level statistical crop yield densities. All continuous physical variables were harmonized, cropped across the CONUS bounding box (25.9°N–49.0°N, 123.1°W–69.1°W), and extracted around target field centroids across three spatial receptive fields ($8\times8$, $24\times24$, and $56\times56$ grids at 0.1° base resolution). Features were resampled to a standardized $8\times8$ patch tensor using bilinear interpolation, while discrete crop masks were preserved using nearest-neighbor interpolation.The dataset spans 18 consecutive production years from 2008 to 2025 with an 8-month monthly temporal sequence covering the key crop growing season (March to October). Stored in a standardized Zarr array format compressed with Blosc (LZ4 level 5), the archive contains 420,102 samples covering two major crops (214,355 corn samples under crop ID 0 and 205,747 soybean samples under crop ID 1). Each sample comprises an 8-month, 36-channel dynamic tensor of shape (8, 36, 8, 8), a 9-channel static geographic/topographic prior tensor of shape (9, 8, 8), a corresponding $8\times8$ binary crop mask, geographic centroid coordinates (lat, lon), production year, crop classification ID, and a scalar target yield label in metric tons per hectare (t/ha). Physical variables have been filtered for quality with no invalid NaN or infinite values, while unharvested or inference-only grid instances without survey ground truth are encoded with a standard sentinel label of -1.0. The data can be directly loaded and processed in Python using the official zarr library (version 2.12 or higher, https://zarr.readthedocs.io) or PyTorch spatial dataloaders without complex decompression overhead. crop yield prediction; MODIS; Zarr Minglin Zou PUBLIC https://creativecommons.org/licenses/by-nc/4.0/ dataset Science Data Bank