ACYD: Agricultural Crop Yield Dataset

Weekly, admin-2 level covariates for Argentina, Brazil, and Mexico (1979–2024), designed for crop yield prediction and benchmarking.

Code & processing pipeline: https://github.com/Neehan/amazing-crop-yield-datasets

The same scripts also support USA — only the --country flag changes. USA data is not included in this release but can be reproduced from the GitHub repo.

Files

This repository contains three directories per country (argentina/, brazil/, mexico/):

• raw/ — direct pulls from source datasets (AgERA5, ERA5, NOAA AVHRR/VIIRS, SoilGrids, GADM, government yield data)
• final/ — processed CSVs ready for ML, one file per variable, aggregated to admin-2 units with cropland-weighted averages on boundaries
• gadm/ — GADM administrative boundaries

intermediate/ files are NOT included — they are regenerable from raw/ via the processing scripts in the GitHub repo.

Variables

Weather

Temperature 2m max/min · wind speed 10m · reference evapotranspiration (RefET) · vapor pressure · snow water equivalent · solar radiation · precipitation.

Source: AG-ERA5 (1979–2024), collected daily and averaged weekly. HYDE-3.5 cropland mask applied within admin boundaries per year; values are weighted averages over cropland fraction.

Land Surface

LAI Low (crops, grass, shrubs) · LAI High (forests, woody vegetation) · NDVI.

Source:

• LAI: ERA5 Landscape Reanalysis, weekly average.
• NDVI: NOAA AVHRR Climate Data Record (1982–2013) and NOAA VIIRS Climate Data Record (2014+).

LAI has no missing values. NDVI has missing weeks (especially 1994 and earlier) — kept as NaN.

Soil

Bulk density · CEC · clay · coarse fragments · nitrogen · organic carbon density · organic carbon content · pH (H₂O) · sand · silt.

Source: SoilGrids / ISRIC (static, published 2020). Aggregated at admin-2 (no cropland mask).

Crops Covered

• Argentina (MAGyP): corn, soybean (total + 1st/2nd cycle), sunflower, wheat
• Brazil (IBGE): corn, wheat, soybean, rice, beans, sugarcane, cotton, sunflower, sorghum, oats, barley, rye, triticale, potato, sweet_potato, cassava, tomato, onion, garlic, peanut, tobacco, watermelon, melon, pineapple, castor, jute, flax, ramie, mallow, pea, fava, alfalfa, sugar_cane_forage
• Mexico (SIAP, with _irrigated / _rainfed variants): corn, soybean, wheat, sorghum, sugarcane, tomato, beans, barley

Geography

• Argentina: 495 weather depts / 483 land surface / 500 soil / 181–296 yield depts
• Brazil: admin-2 (município) level
• Mexico: admin-2 (municipio) level

Time Span

• Weather, LAI: 1979–2024
• NDVI: 1982–2024
• Crop yields: 1970–2025 (varies by crop/country)
• Soil: static (2020)

Reproducing

GitHub repo has the full pipeline: download → standardize to NetCDF → spatial aggregation with HYDE cropland mask → admin-2 weekly CSV. Key scripts:

• cli/download_*.py — download raw data
• cli/process_*.py — produce admin-2 CSVs
• process_all_data.sh — end-to-end pipeline

Licenses

Preprocessing code is MIT. Underlying data retains original licenses:

• ERA5 / AgERA5: Copernicus (CC BY 4.0 equivalent)
• NOAA AVHRR / VIIRS NDVI: Public Domain (U.S. Government work)
• HYDE 3.5 Cropland: CC BY 4.0
• IBGE (Brazil yields): CC BY 3.0
• MAGyP (Argentina yields): CC BY 4.0
• SIAP (Mexico yields): public, check source
• FAO (if included): CC BY-NC-SA 3.0