Weakly supervised framework for wildlife detection and counting in challenging Arctic environments: a case study on caribou (Rangifer tarandus)

1. Problem Statement

The study addresses the critical challenge of monitoring declining caribou populations in the Arctic using aerial imagery. Traditional manual interpretation is labor-intensive, error-prone, and unscalable. While Deep Learning (DL) offers automation, it faces significant hurdles in this specific domain:

• Severe Class Imbalance: Aerial mosaics are dominated by empty terrain, with animals occupying a tiny fraction of the image, leading to high false-negative rates.
• Background Heterogeneity: The Arctic landscape includes diverse land covers (tundra, wetlands, snow, shrublands) with varying textures and colors, causing animals to blend in or be confused with background features (false positives).
• Target Characteristics: Caribou appear as small, often occluded objects that vary significantly in density (from sparse individuals to dense herds of thousands) and scale.
• Domain Mismatch: Standard initialization using ImageNet weights (trained on ground-level, proximal objects) often fails to transfer effectively to high-altitude, nadir-view wildlife detection, particularly when animals are sparse or visually subtle.
• Data Scarcity: Annotating large-scale aerial imagery is prohibitively expensive and time-consuming, limiting the availability of fully supervised datasets.

2. Methodology

The authors propose a two-stage weakly supervised framework designed to learn dataset-specific features before performing the main detection task. The pipeline consists of:

A. Data Preparation

• Dataset: Aerial imagery from five caribou herds in Alaska (CAH, FCH, PCH, TCH, WAH) acquired in 2017 and 2019.
• Patching: Large images were decomposed into 512 × 512 pixel patches with a 15% overlap to ensure animals near boundaries are captured.
• Weak Labels: Instead of bounding boxes for every animal, patches were labeled with coarse binary labels: Empty (0) or Non-Empty (1).
• Class Balancing: A Binary Batch Sampler (BBS) was used during training to ensure equal numbers of empty and non-empty patches in every mini-batch, mitigating the natural 5:1 imbalance.

B. Stage 1: Patch-based Pretrain Network (PPN)

• Architecture: A HerdNet backbone (DLA-34 encoder-decoder) was repurposed as a binary classifier.
• Task: Classify patches as "empty" or "non-empty."
• Initialization Strategies:
  1. PPN-ImageNet: Initialized with ImageNet weights.
  2. PPN-Scratch: Initialized with random weights.
• Loss Function: Binary Cross-Entropy (BCE).
• Goal: To learn features specific to the Arctic environment and the presence of caribou, creating a "weakly supervised" knowledge base.

C. Stage 2: Point-based Detection Network (HerdNet)

• Architecture: The original HerdNet (a point-based detector using Focal Instance Distance Transform heatmaps).
• Initialization: The backbone weights were transferred from the trained PPN models. Three scenarios were tested:
  1. HerdNet-PPN-ImageNet: Backbone weights from PPN (trained on ImageNet init).
  2. HerdNet-PPN-Scratch: Backbone weights from PPN (trained from scratch).
  3. HerdNet-ImageNet (Baseline): Backbone weights directly from ImageNet (standard approach).
• Hard Negative Mining (HNM): After initial training on positive patches, Hard Negative Patches (HNPs) (background patches previously misclassified as containing animals) were extracted and added to the training set to reduce false positives.

3. Key Contributions

1. Weakly Supervised Pretraining for Domain Adaptation: The paper introduces a novel approach where a detection network's backbone is first pretrained on the same target dataset using coarse patch-level labels (empty vs. non-empty). This bridges the domain gap better than generic ImageNet pretraining.
2. Demonstration of "From Scratch" Viability: The study shows that even without ImageNet initialization, pretraining on the target dataset (PPN-Scratch) yields features that outperform or match generic ImageNet initialization, proving that dataset-specific feature learning is crucial.
3. Robustness Across Years and Herds: The framework was validated on independent data from a different year (2019) and multiple distinct herds, demonstrating strong generalization capabilities.
4. Efficiency in Annotation: The PPN acts as a pre-filter, allowing survey teams to discard empty regions automatically, significantly reducing the manual workload for human reviewers.

4. Results

The models were evaluated on positive patches and full-image detection/counting tasks.

• PPN Performance (Patch Classification):

  • PPN-ImageNet achieved the highest accuracy: F1-score of 93.7% (2017) and 92.6% (2019).
  • PPN-Scratch performed significantly worse (F1 ~70%), indicating that random initialization struggles to learn meaningful features on limited data without the benefit of generic pretraining.

• Detection Performance (Positive Patches):

  • HerdNet-PPN-ImageNet outperformed all other models:
    • F1-score: 92.6% (2017) and 93.5% (2019).
    • Recall: 92.7% and 96.1% (significantly higher than the baseline).
    • Mean Absolute Error (MAE): 0.8 and 0.7 (lowest error).
  • The baseline HerdNet-ImageNet consistently undercounted animals (negative Total Counting Error).

• Full-Image Performance (with Hard Negatives):

  • HerdNet-PPN-ImageNet achieved the best overall full-image metrics:
    • F1-score: 95.5% (2017) and 93.3% (2019).
    • Compared to the baseline (91.5% / 90.4%), this represents a substantial improvement in both precision and recall.
  • HerdNet-PPN-Scratch also showed competitive results, outperforming the baseline in recall, proving that the weakly supervised pretraining step itself is the primary driver of performance gains.

5. Significance and Implications

• Overcoming Data Scarcity: The study proves that effective wildlife detection is possible even with limited annotated data by leveraging weak supervision (patch-level presence) to adapt models to specific environmental conditions.
• Scalability for Conservation: The framework enables scalable monitoring of vast Arctic regions. By filtering out empty patches, it drastically reduces the time required for manual verification, making large-scale population estimates feasible.
• Generalizability: The success across different herds and years suggests this approach can be applied to other circumpolar regions (Canada, Scandinavia, Russia) and potentially other wildlife species with similar detection challenges.
• Future Directions: The authors suggest extending this to multi-class detection (e.g., distinguishing calves from adults) and using the trained model to generate pseudo-labels for unannotated data, further reducing the need for manual labeling.

In conclusion, the paper demonstrates that pretraining a detection backbone on the target dataset using weak, coarse labels is a superior strategy to generic ImageNet initialization for complex, imbalanced, and heterogeneous aerial wildlife monitoring tasks.