README.md

---
license: cc-by-4.0
tags:
  - remote-sensing
  - satellite-imagery
  - agriculture
  - classification
  - multi-label
  - bounding-box
dataset_type: image-classification
pretty_name: CAFOSat Dataset
size_categories:
  - 10K<n<100K
---

# CAFOSat: CAFO Infrastructure Dataset

CAFOSat is a remote sensing dataset designed for identifying and classifying Concentrated Animal Feeding Operations (CAFOs) across various U.S. states. It includes high-resolution image patches, infrastructure annotations, bounding boxes, and experimental train-test splits for multiple configurations.

---

## Dataset Structure

The dataset is organized into extracted folders originally packaged as `.tar.gz` files:

- `STATE_filtered/`: Original patches per U.S. state  
  (e.g., `IA_filtered/`, `AL_filtered/`)
- `negative_samples/`: Verified non-CAFO examples
- `barn/`, `manure_pond/`, `others/`: Augmented synthetic patches by infrastructure type
- `cafosat.csv`: Master metadata file with labels, bounding boxes, and split flags

All image paths referenced in the CSV point to these extracted folders.

---

## 🔖 Image File Reference (`patch_file`)

Each row in the metadata includes a `patch_file` field that provides the relative path to the associated image file.
Each `patch_file` is a pointer into a compressed archive using Hugging Face's streaming format:

Example:  
-`IA_filtered.tar.gz::IA_filtered/crop_4517_patch_10147_Swine_Nursery_IA.tif`
- `barn.tar.gz::`barn/aug_patch_00123.tif`
- `negative_sample.tar.gz::`barn/neg_patch_00098.tif`

This format indicates the image is located inside `IA_filtered.tar.gz` under the subpath shown. This field is automatically interpreted by Hugging Face as an image using the `datasets.Image()` feature, so image previews and loading work out of the box.

---

## Features

| Column        | Description                                           |
|---------------|-------------------------------------------------------|
| `patch_file`  | Path to the image file                                |
| `label`       | Integer label for class (0–6)                         |
| `barn`, `manure_pond`, `grazing_area`, `others` | Binary infra flags       |
| `geom_bbox`   | Bounding box coordinates `[x1, y1, x2, y2]`           |
| `category`    | Class name (e.g., Swine, Dairy)                       |
| `state`       | U.S. state of the patch                               |
| `verified_label` | Human-verified CAFO type                           |
| `CAFO_UNIQUE_ID` | Unique identifier for facility                     |
| `image_type`  | `original`, `augmented`, `negative`, etc.             |
| `split columns` | Flags for different train/test/val splits           |

---

## Labels

| Class ID | Class Name     |
|----------|----------------|
| 0        | Negative        |
| 1        | Swine           |
| 2        | Dairy           |
| 3        | Beef            |
| 4        | Poultry         |
| 5        | Horses          |
| 6        | Sheep/Goats     |

---

## Splits

Multiple experimental train-test split columns are provided in the CSV:

- `cafosat_verified_training_train`, `cafosat_verified_training_test`, `cafosat_verified_training_val`
- `cafosat_all_training_*`
- `cafosat_training_set1_*`, `set2_*`
- `cafosat_merged_training_*`
- `cafosat_augmented_training_*`

Each flag is a binary indicator (`1` = in split, `0` = excluded).

---

## Intended Use

- CAFO detection and classification
- Agricultural infrastructure mapping
- Weak supervision, semi-supervised learning
- Remote sensing benchmark development

---

## Citation

*TBD

---

## License

This dataset is released under the [CC BY 4.0 License](https://creativecommons.org/licenses/by/4.0/).

You are free to share, use, and adapt the data with attribution.


## Contact

For questions or contributions, contact `[email protected]`.