Generating Synthetic Wildlife Health Data from Camera Trap Imagery: A Pipeline for Alopecia and Body Condition Training Data

1. Problem Statement

The Data Gap in Wildlife Health Monitoring:
While camera trap networks generate tens of millions of images annually and computer vision has successfully automated species identification and counting, automated health screening remains absent.

• Current Limitation: Health observations (e.g., mange, emaciation) are currently made opportunistically by human biologists, are not systematically recorded, and do not scale.
• Root Cause: The barrier is a lack of publicly available, ML-ready datasets for wildlife health conditions. Existing datasets are either manually labeled (small scale) or synthetic but focused only on species classification, not health phenotypes.
• Diagnostic Complexity: Visual assessment alone cannot provide a clinical diagnosis (e.g., distinguishing mange from lice or seasonal shedding). Therefore, the goal is not a definitive diagnosis but a "suspect flag" system to prioritize images for expert review.

2. Methodology

The authors propose a three-stage pipeline to generate synthetic training data from real camera trap imagery, explicitly designed for screening rather than clinical diagnosis.

Stage 1: Base Image Curation

• Source: Images are curated from iWildCam 2022.
• Target Species: 8 North American species (White-tailed deer, Mule deer, Elk, Gray wolf, Red fox, Gray fox, Raccoon, and a broad cervid category).
• Selection Strategy:
  • Uses MegaDetector v4 to identify animal bounding boxes.
  • Employs center-frame-weighted stratified sampling (prioritizing animals in the center of the frame) to ensure consistent framing.
  • Maintains proportional species representation and temporal diversity (seasons, day/night).
• Dataset: A balanced base set of 1,000 images was constructed, from which a stratified subsample of 201 images was used for the study.

Stage 2: Phenotype Editing (Generative Model)

• Model: Uses Gemini 3.1 Flash Image to edit base images.
• Phenotypes Generated: Two visual dimensions with three severity levels each:
  1. Alopecia (Hair Loss): M0 (Healthy), M2 (Moderate patchy loss/scaling), M3 (Severe loss/hyperkeratosis).
  2. Body Condition: B0 (Healthy), B2 (Visible ribs/wasting), B3 (Severe emaciation).
• Variant Design: Each base image generates up to four variants:
  • Sham (M0/B0): Negative control (no edit).
  • Alopecia-only (M2/B0).
  • Emaciated-only (M0/B2).
  • Severe-both (M3/B3).
• Biological Constraints: Prompts include species-specific lesion distributions (e.g., face/ears for canid mange; spine-sparing for hoofstock due to grooming anatomy).
• Spatial Constraints: Edits are strictly anchored to the MegaDetector bounding box to preserve the background, lighting, and camera artifacts pixel-identically.

Stage 3: Scene-Drift Quality Control (QC)

To ensure the generative model did not hallucinate or alter the background (scene drift), a rigorous QC system is applied:

• Sham Pre-filter: The "Sham" variant is generated first. If it fails QC (indicating the model fundamentally re-rendered the scene), all other variants for that base image are skipped, saving API costs.
• Mask Construction: A mask is built from raw pixel differences, filtered to only include changes overlapping the animal's bounding box.
• Scoring Metrics: Quality is scored on blurred pixels outside the animal mask using three metrics:
  1. Raw Mean Absolute Error (MAE).
  2. Color-normalized MAE (correcting for color shifts).
  3. Structural Similarity Index (SSIM).
• Adaptive Thresholds:
  • Daytime: Passes if Normalized MAE ≤ 7.0 OR SSIM ≥ 0.85.
  • Nighttime/IR: Stricter threshold (Raw MAE ≤ 5.0) due to minimal background variation in IR.

3. Key Contributions

1. First Pipeline for Synthetic Wildlife Health Data: The first system to generate synthetic training data for wildlife health conditions (alopecia and body condition) from camera trap imagery, featuring adaptive quality control.
2. Biologically Informed Phenotype Design: Incorporates species-specific lesion distributions and explicitly frames outputs as "screening data" (suspect flags) rather than clinical diagnoses.
3. Sim-to-Real Transfer Validation: Demonstrates that a classifier trained exclusively on synthetic data can effectively detect suspected health conditions in real-world camera trap images.

4. Results

• Data Generation Efficiency:
  • Processed 201 base images across 4 species.
  • Generated 666 total variants.
  • 553 variants (83%) passed QC.
  • The Sham pre-filter rejected 16% of base images early, saving 93 unnecessary API calls.
• QC Performance:
  • Pass rates were consistent between Day (85%) and Night (81%).
  • The "Severe-both" variant had the lowest pass rate (82%), confirming that larger edits induce more scene drift.
• Sim-to-Real Transfer Experiment:
  • Setup: Trained a classifier (DINOv2 backbone + MLP head) only on the 553 synthetic images. Tested on 70 real images (45 healthy, 25 suspected health issues).
  • Performance:
    • MLP Model: Achieved 0.854 AUROC and 0.811 balanced accuracy.
    • Linear Probe: Achieved 0.734 AUROC (suggesting the visual signal is non-linear).
  • Conclusion: The synthetic data captures sufficient visual features to distinguish healthy from suspect animals in real-world scenarios without any real training data.

5. Significance and Limitations

Significance:

• Overcomes Data Scarcity: Solves the fundamental barrier of lacking labeled health datasets for wildlife.
• Scalable Screening: Enables the deployment of automated systems to flag potential health crises in massive camera trap datasets, allowing biologists to focus on high-priority images.
• Feasibility Proof: Proves that synthetic data can bridge the domain gap to real-world wildlife health monitoring.

Limitations:

• Diagnostic Ambiguity: The system flags "hair loss consistent with mange" but cannot distinguish between mange, lice, or seasonal shedding.
• Species-Condition Mismatch: In some regions, mange in white-tailed deer is rare; the model may flag seasonal shedding as a false positive, requiring regional calibration.
• Posture Confounds: The current pipeline edits texture and body shape but does not modify animal posture, which is a key indicator of illness.
• Evaluation Scale: The real-world test set is small (N=70) and relies on community-reported labels rather than clinical confirmation.

Future Work:
The authors plan to validate synthetic realism with wildlife pathologists, expand the evaluation with clinical cases, and integrate pose-conditioned generative models to capture illness-related posture changes. Preliminary deployment on 22,958 images already flagged ~5% for expert review.