reglab/wisconsin-dairy-cafo

Wisconsin Large Dairy Farms Dataset (CAFOs)

This is a geo-referenced dataset of large dairy farms in Wisconsin compiled using satellite imagery, computer vision, human validation, and merges to key administrative and open geospatial data sources.

The process to generate this data is described in the manuscript: Blind Spots in Environmental Permitting: An AI-Assisted Assessment of Coverage Gaps and Environmental Risk among Large Dairy Farms in Wisconsin (link to be made available upon publication).

We also include the input datasets which, along with the analysis code at this GitHub repository, can reproduce the paper's key results.

---

Repository Structure

wi_cafo_facilities.geojson       ← Primary publication dataset (WGS84)
wi_cafo_facilities.parquet       ← Same data in GeoParquet (WI State Plane EPSG:3071)
wi_cafo_facilities.csv           ← Same data with geometry as WKT column
input_data/
    annotations/
        full_state_cf_annotations.geojson
        ewg_region_train_labels.geojson
    WDNR_CAFOs.geojson
    ewg_AFOs_012022.geojson
    permit_animal_type_records.csv
    milk_producers.geojson
    clusters/
        all_cf_clusters.geojson
        four_band_clusters.csv
        three_band_clusters.csv
    geospatial/
        county_boundaries/
        water_data/

---

Publication Dataset (wi_cafo_facilities.*)

The main output of the study: 1,469 large dairy farms across Wisconsin, with permit linkage, animal unit estimates, milk producer license linkage, and environmental risk indices.

The process to generate this data is described in the manuscript (to be linked upon publication). It includes data on:

• All permitted dairy CAFO farms as per 2022 WPDES permit data from Wisconsin Department of Natural Resources
• Dairy farms found by our model and human validation pipeline, estimated to contain 500 animal units or more, that did not match to 2022 WPDES permits.

Column Descriptions

Column	Type	Source	Description
polygon_indices	int	Constructed	Unique serial identifier for each farm, zero-indexed.
animal_unit_estimate	float	Constructed	Point estimated animal units (AU) at the farm. Method described in manuscript: mean value of a Monte Carlo simulation using total barn area, space-per-animal and cow stage distributions.
animal_units_lower	float	Constructed	Lower bound estimated animal units. Calculated as the 2.5th percentile of the simulated uncertainty interval method.
animal_units_upper	float	Constructed	Upper bound of the 95% simulation interval for the AU estimate. Calculated as the 97.5th percentile of the simulated uncertainty interval method.
type	string	Constructed	Facility classification based on permit matching and estimated animal units. Categories: "Permitted dairy CAFOs" (matched to a WDNR 2022 WPDES CAFO permit), "AFOs between 500 and 1000 AU" (estimated 500–1000 AU, no permit match), "Potential unpermitted CAFOs (>1000 AU)" (estimated >1000 AU, no permit match — above the threshold requiring a CAFO permit in Wisconsin).
permit_facility_id	float	WDNR permit data	Wisconsin DNR Facility Identification Number (FIN) for facilities matched to a CAFO permit within 500m. NaN for unmatched facilities. Source: WDNR_CAFOs.geojson.
permit_allowable_animal_units	float	WDNR permit data	Maximum animal units allowed under the facility's CAFO permit. NaN for unmatched facilities. Source: WDNR_CAFOs.geojson.
matched_milk	bool	WI DATCP milk producer licenses	Whether the facility was spatially matched (within 500m) to an active Wisconsin milk producer license. Source: milk_producers.geojson.
type_of_milk	string	WI DATCP milk producer licenses	Type of milk produced at the matched milk license ("Bovine", "Goat", etc.). None if matched_milk is False. Source: milk_producers.geojson.
water_distance	float	WI DNR hydrography	Distance (m) from the facility centroid to the nearest surface water body (open water, perennial or intermittent stream).
impaired_water_distance	float	EPA 303(d) / WI DNR	Distance (m) to the nearest EPA 303(d)-listed impaired water body or stream.
closest_water_impaired	bool	EPA 303(d) / WI DNR	True if the nearest water body is 303(d)-listed impaired (i.e., impaired_water_distance ≤ water_distance).
hydro_intermit_distance	float	WI DNR 24K hydrography	Distance (m) to the nearest intermittent stream.
hydro_perennial_distance	float	WI DNR 24K hydrography	Distance (m) to the nearest perennial stream.
gw_0	bool	WI DNR groundwater	True if the facility is located in an area with a water table depth of 0 ft (i.e., at or above the surface). Source: GCSM water table depth layer.
gw_20	bool	WI DNR groundwater	True if the facility is in an area with water table depth ≤ 20 ft.
gw_50	bool	WI DNR groundwater	True if the facility is in an area with water table depth ≤ 50 ft.
bedrock_lt5ft_distance	float	SNAPMAPs	Distance (m) to the nearest area with bedrock depth < 5 ft (shallow bedrock). Source: SNAPMAPs soils layer.
bedrock_lt5ft_dummy	bool	SNAPMAPs	True if the facility itself is located on shallow bedrock (depth < 5 ft).
silurian_0_2_distance	float	SNAPMAPs	Distance (m) to the nearest area with Silurian carbonate bedrock at 0–2 ft depth.
silurian_0_2_dummy	bool	SNAPMAPs	True if the facility sits on Silurian carbonate bedrock at 0–2 ft.
silurian_2_5_distance	float	SNAPMAPs	Distance (m) to the nearest area with Silurian carbonate bedrock at 2–5 ft depth.
silurian_2_5_dummy	bool	SNAPMAPs	True if the facility sits on Silurian carbonate bedrock at 2–5 ft.
shallow_silurian_distance	float	SNAPMAPs	Distance (m) to the nearest area with any shallow Silurian carbonate bedrock (0–5 ft).
shallow_silurian_dummy	bool	SNAPMAPs	True if the facility sits on any shallow Silurian carbonate bedrock (0–5 ft).
mean_slope	float	WI DNR 30m DEM	Mean terrain slope (%) within a 1 km radius of the facility centroid, derived from a 30m digital elevation model.
slope_greater_12_distance	float	WI DNR 30m DEM	Distance (m) to the nearest area with slope > 12%.
slope_greater_12_dummy	bool	WI DNR 30m DEM	True if the facility is located in an area with slope > 12%.
swqma_300ft_distance	float	SNAPMAPs	Distance (m) to the nearest Surface Water Quality Management Area (SWQMA) 300 ft buffer zone — a regulatory setback for manure spreading.
swqma_300ft_dummy	bool	SNAPMAPs	True if the facility is within a SWQMA 300 ft zone.
swqma_1000ft_distance	float	SNAPMAPs	Distance (m) to the nearest SWQMA 1,000 ft buffer zone.
swqma_1000ft_dummy	bool	SNAPMAPs	True if the facility is within a SWQMA 1,000 ft zone.
cafo_w_restrict_distance	float	SNAPMAPs	Distance (m) to the nearest winter-restricted spreading zone for CAFOs (areas where manure spreading is prohibited in winter).
cafo_r_restrict_distance	float	SNAPMAPs	Distance (m) to the nearest CAFO regulatory restriction area (general CAFO-specific spreading restriction zone).
hand_built_risk_index	float	Constructed	Composite environmental risk index (0–1) built by manually weighting proximity to impaired waters, groundwater depth, bedrock depth, and regulatory restriction zones. Higher = greater environmental risk.
pca_risk_index	float	Constructed	Composite environmental risk index (0–1) derived from principal component analysis of the same geospatial risk factors used in hand_built_risk_index. Higher = greater environmental risk.
n_buildings	int	Constructed	Number of individual barn polygons that make up this facility cluster.
n_parcels	int	Constructed	Number of distinct land parcels associated with this facility cluster (via spatial join to WI statewide parcel data).
cluster_area_m2	float	Constructed	Total footprint area of all barn polygons in the facility cluster, in square meters.
geometry	geometry	Constructed	MultiPolygon of all human-annotated barn footprints comprising the facility, in WGS84 (EPSG:4326) for GeoJSON/CSV and WI State Plane (EPSG:3071) for Parquet.

Quick Start

import geopandas as gpd

# Load from local file
gdf = gpd.read_file("wi_cafo_facilities.geojson")

# Or load from Hugging Face Hub
from huggingface_hub import hf_hub_download
path = hf_hub_download(
    repo_id="reglab/wisconsin-dairy-cafo",
    filename="wi_cafo_facilities.geojson",
    repo_type="dataset"
)
gdf = gpd.read_file(path)

---

Input Data (input_data/)

These are the raw and intermediate datasets used to generate the publication dataset and all paper figures. Download this folder and follow the setup instructions at this GitHub repository to reproduce all results.

input_data/annotations/

Human-labeled barn polygon annotations produced by CloudFactory workers on NAIP aerial imagery.

File	Description
full_state_cf_annotations.geojson	All human-annotated polygon annotations outside the 9-county EWG study region.
ewg_region_train_labels.geojson	All human-annotated polygon annotations within the 9-county EWG study region (Brown, Calumet, Fond du Lac, Green Lake, Manitowoc, Outagamie, Sheboygan, Waupaca, Winnebago).

---

input_data/WDNR_CAFOs.geojson

Source: Wisconsin Department of Natural Resources (WDNR) Acquired: 2022-01-25

Locations and full permit metadata for all CAFO sites permitted by WDNR in Wisconsin. This is a full join of CAFO main permit sites (2018 database) and satellite sites (2022 geocoded permits).

---

input_data/ewg_AFOs_012022.geojson

Source: Environmental Working Group (EWG) Acquired: January 2022

Point locations and metadata for all Animal Feeding Operations (AFOs) labeled by EWG across Wisconsin. See description and methodology here

---

input_data/permit_animal_type_records.csv

Source: WDNR permit website (manual compilation) Date: 2025-05-14

Animal type counts manually compiled from WDNR CAFO permit application PDF documents. Each row corresponds to one permitted facility and contains counts of each animal type on-site (dairy cows, heifers, calves, dry cows, swine, poultry, etc.). Sourced from the WDNR e-permitting portal website here.

Key columns:

• Facility identifiers linking to WDNR_CAFOs.geojson via FIN
• Per-animal-type counts (dairy cows, heifers, calves, dry cows, beef, swine, poultry)

---

input_data/milk_producers.geojson

Source: Wisconsin Department of Agriculture, Trade and Consumer Protection (DATCP) Acquired: 2023-07-01, updated 2024-03-04 URL: DATCP Milk Producer Licenses

All active permitted milk producer licenses in Wisconsin, geocoded to coordinate points using the Google Geocoding API. 75 instances of inaccurate geocoding were identified and manually corrected in the March 2024 update.

---

input_data/clusters/

Facility-level clusters: groups of overlapping/adjacent model-predicted barn polygons (or human-annotated polygons) merged into single facility records using the clustering algorithm in cluster.py. All model clusters use confidence threshold 0.5 unless otherwise noted.

File	Size	Description
all_cf_clusters.geojson	~14 MB	Facility clusters derived from all CloudFactory human annotations statewide. Includes cluster geometries.
four_band_clusters.csv	~224 MB	Facility clusters from the 4-band (RGB + near-infrared) segmentation model, run at full state coverage.
three_band_clusters.csv	~180 MB	Facility clusters from the 3-band (RGB only) segmentation model, run at full state coverage.

---

input_data/geospatial/

Ancillary geospatial layers used in environmental risk assessment.

geospatial/county_boundaries/

Wisconsin county boundary shapefiles at 1:24,000 scale. Source: WDNR Open Data Portal

geospatial/water_data/

File	Source	Description
GCSM_-_Water_Table_Depth.geojson	WI DNR	Depth to groundwater table (m) across Wisconsin
impaired_rivers_streams.geojson	EPA 303(d) / WI DNR	Rivers and streams listed as impaired under the Clean Water Act
impaired_lakes.geojson	EPA 303(d) / WI DNR	Lakes listed as impaired under the Clean Water Act
24k_Hydro_Waterbodies_(Open_Water).geojson	WI DNR	All open water bodies in Wisconsin (24K hydrography)

---

Not Included (Available from Original Sources)

These large datasets are used in the analysis pipeline but are not included here. Download them from the original sources and set paths in config/config.yml per the analysis repo README.

Dataset	Source	Notes
Wisconsin land parcels (V8.0.0)	WI SCO Parcel Data	Full statewide parcel polygons; used for ownership matching and clustering of individual barn polygons into farms
30m Digital Elevation Model	WI DNR / geodata.wisc.edu	Used to compute mean slope within 1km radius of each facility
Wisconsin urban areas (TIGER 2020)	U.S. Census TIGER	Used to filter model predictions to exclude urban areas
SNAPMAPs nutrient management layers	DATCP GIS Data	Policy-relevant soils, waterways, slope, and regulatory restriction layers used in risk assessment
USDA NAIP aerial imagery	NAIP Hub	Used for model training and inference.

---

Reproducing the Paper

1. Clone reglab/wi-cafo-analysis and follow its README to set up the Python environment.
2. Download input_data/ from this repository and place it at the paths specified in config/config.yml.
3. Download the "Not Included" datasets above from their original sources.
4. Run:

   python generate_paper_results.py --skip-imagery  # omit --skip-imagery if you have GCP access
   

5. All paper figures and tables will be written to paper_results/.

---

License

This dataset is released under CC BY 4.0. You are free to share and adapt for any purpose with attribution.

Source data (WDNR permits, EWG AFOs, DATCP milk licenses, WI DNR hydrography) are from public government sources.

---

Citation

Citation to be added upon publication.

---

Contact

RegLab, Stanford Law School — reglab.stanford.edu