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

Technical Summary: CAFOSat

Problem Statement
Concentrated Animal Feeding Operations (CAFOs) are critical nodes for disease surveillance and climate resilience planning but pose significant environmental and public health risks. Despite their importance, scalable mapping of CAFOs remains hindered by fragmented data sources, noisy point locations (often referring to parcel centroids or access roads rather than facility footprints), inconsistent annotations, and systematic under-reporting. Existing datasets are often geographically limited, lack infrastructure-level detail (e.g., barns, manure ponds), and suffer from poor generalization due to the visual variability of CAFOs across different livestock types and regions. Furthermore, the labor-intensive nature of manually refining weak geolocations to align with high-resolution imagery creates a bottleneck for creating high-quality training datasets.

Methodology
The authors introduce CAFOSat, a strongly annotated, infrastructure-aware dataset covering 20 U.S. states. The construction of CAFOSat follows a five-step pipeline designed to transform weak, publicly available inventories into high-fidelity training data:

1. Data Aggregation & Standardization: The process begins by aggregating CAFO location inventories from federal, state, and academic sources (e.g., CAFOMaps). Initial coordinates are mapped to high-resolution NAIP imagery (0.6m resolution), and fixed-size 833×833 pixel patches are extracted.
2. AI Annotator & Pre-filtering: To reduce manual effort, a ResNet-based binary classifier is trained on a small set of manually verified samples (from RegLab). This "AI annotator" pre-filters candidate patches from the broader inventory, retaining only those with high-confidence predictions (softmax ≥ 0.8) for human review.
3. Coordinate Refinement (Weak-to-Strong): A novel refinement algorithm converts noisy coordinates into spatially accurate point annotations. This involves:
   • Extracting five overlapping patches (one centered, four diagonally offset) around the weak coordinate.
   • Applying GradCAM to generate class-specific attention heatmaps for patches predicted as CAFOs.
   • Thresholding heatmaps to isolate activation regions, converting them to polygons, and applying single-linkage clustering to group contiguous structures.
   • Computing the centroid of the largest cluster as the refined coordinate.
4. Manual Verification & Infrastructure Annotation: A total of 4,513 samples are manually verified. Beyond binary classification, annotators provide infrastructure-level labels, including counts of barns and manure ponds, and the presence of grazing areas or other structures.
5. Curated Negatives & Synthetic Augmentation:
   • Hard Negatives: Non-CAFO samples are generated using land-cover-guided sampling (stratified across pasture, cropland, etc.) with a 1.5km exclusion buffer around known CAFOs to ensure they are true negatives.
   • Synthetic Augmentation: A prompt-guided pipeline uses GroundingDINO for open-vocabulary infrastructure detection and Stable Diffusion for inpainting. This removes specific structures (e.g., barns) and replaces them with semantically plausible background content (e.g., grassland), creating label-preserving variations to diversify the training distribution without altering the CAFO type label.

Key Contributions

• Benchmark Dataset: CAFOSat comprises over 45,000 image patches (39,257 base + 6,454 synthetic) spanning 20 states and four major CAFO types (Swine, Poultry, Dairy, Beef). It includes unique infrastructure-level annotations and a curated set of hard negatives.
• Scalable Annotation Pipeline: The authors propose a human-in-the-loop framework that significantly reduces manual annotation effort while improving label quality through AI pre-filtering and GradCAM-based refinement.
• Comprehensive Evaluation: The paper provides a rigorous benchmarking suite evaluating modern CNNs, Vision Transformers (ViT, Swin), and Vision-Language models (CLIP, RemoteCLIP) across in-distribution and out-of-distribution (unseen states) settings.
• Reproducibility: All code, data processing scripts, annotation tools, and data loaders are publicly released to facilitate adoption.

Results

• Impact of Annotation Quality: Experiments demonstrate that refining weak coordinates to align with actual infrastructure significantly improves classification performance. Models trained on refined patches consistently outperform those trained on weak survey coordinates across all architectures.
• Model Performance: Transformer-based models (e.g., Swin-B, EfficientNet) generally outperform classical CNNs (ResNet) in terms of accuracy, F1-score, and mean Average Precision (mAP). Poultry is identified as the easiest class to recognize, while Dairy is the most challenging due to visual similarity with Beef facilities.
• Cross-Dataset Generalization: Models trained on CAFOSat generalize well to external datasets (METER-ML, RegLab) and unseen states, whereas models trained on smaller, less diverse datasets struggle when transferred to CAFOSat.
• Synthetic Augmentation: Training on the synthetic augmented set yields performance comparable to training on manually verified data (a drop of only ~0.6% F1 and ~2.2% mAP for Swin-B), suggesting that targeted inpainting can effectively diversify training data and improve robustness, particularly for transformer-based models.

Significance
The paper posits that the misalignment between public inventory coordinates and actual facility footprints is a primary bottleneck in CAFO mapping. CAFOSat addresses this by demonstrating that refining weak point labels is a first-order driver of performance, often rivaling the impact of changing model architectures. By providing a dataset with infrastructure-level granularity and robust negative sampling, CAFOSat enables more accurate risk assessment and epidemiological modeling. The work highlights that global context and layout cues (captured well by transformers) are critical for distinguishing CAFOs from visually similar agricultural facilities. While the dataset is U.S.-centric and inherits biases from public inventories, it establishes a foundation for scalable, infrastructure-aware livestock mapping and future work in detection, segmentation, and temporal monitoring.