CAFOSat: A Strongly Annotated Dataset for Infrastructure-Aware CAFO Mapping Using High-Resolution Imagery

🗺️ CAFOSAT: Dataset Summary

🔗 CAFOSat is hosted on the Hugging Face Hub: https://huggingface.co/datasets/oishee3003/CAFOSat

🔍 Explore the dataset, its features, and an example dataloader: CAFOSat/data_loader

CAFOSat is a large-scale, high-resolution dataset designed to support infrastructure-aware mapping of Concentrated Animal Feeding Operations (CAFOs) across the United States. The dataset spans 2,064 counties and contains over 45,000 image patches extracted from NAIP imagery, each at 833×833 pixels with a 0.3-0.6m spatial resolution.

📦 Key Features

• Diverse Coverage: Images span six major CAFO types and are sourced from a combination of federal, state, and academic datasets, ensuring diverse geographic and operational representation.
• High-Quality Annotations:
  • Facility-Level Labels: Each image patch includes a CAFO classification label and a bounding box localizing the CAFO infrastructure within the patch.
  • Infrastructure-Level Labels: Detailed manual annotations for structures such as barns, manure lagoons, feed storage, and more for ~3000 CAFO Location.
• Synthetic Subset: Includes a manually validated set of synthetic images to support testing of domain generalization and augmentation robustness. (~6500 Patches)
• Hard Negatives: Carefully curated negative samples that closely resemble CAFO environments but do not contain actual facilities — ideal for training discriminative models.(~23k Patches)

📁 Dataset Structure

CAFOSat/
├── {state_filtered}
│   └── <patch_file>.tif
├── barn/
│   └── <patch_file>.tif
├── manure_pond/
│   └── <patch_file>.tif
├── others/
│   └── <patch_file>.tif
├── negative_sample/
│   └── <patch_file>.tif
├── CAFOSat.csv

📝 Description

• {state}_filtered/: Original geolocation image patches organized by state, including refined geolocation image patches organized by state.
• barn/manure_pond/others: Synthetic data generated by augmentation techniques, categorized by infrastructure type.
• negative_sample/: Hard negative samples—image patches visually similar to CAFOs but containing no actual CAFOs.
• cafosat.csv: CSV files containing metadata for both real and augmented patches (e.g., geolocation, type, annotations, bounding box, weak geolocation, refined geolocation etc).

🛠️ Create Your Own Dataset or Reproduce CAFOSat

To create your own dataset using the CAFOSat annotation and processing pipeline, follow the guide in the repository below:

👉 CAFOSat Data Processing Pipeline

This includes:

• Geolocation Refinement
• Patch generation
• Manual Verfication Setup
• Annotate Infrastructure
• Synthetic sample creation

🧠 Train your own classifier: CAFOSat/classifier