Deep Learning-Based Animal Footprint Analysis for Species Recognition and Injury Detection

Technical Summary: Deep Learning-Based Animal Footprint Analysis for Species Recognition and Injury Detection

Problem Statement
Wildlife management and ecosystem monitoring rely heavily on the detection and tracking of injured animals. Traditional methods involve manual footprint analysis (ichnology), which is inherently subjective, time-consuming, and difficult to scale due to the need for extensive domain expertise. Furthermore, environmental confounding factors—such as varying substrate types (e.g., sand vs. clay), lighting conditions, and incomplete prints—introduce high intra-class variance, distorting track morphology. While morphological features like toe geometry and pad configurations offer diagnostic value for species identification and injury detection (e.g., lameness or trauma), current automated systems remain underutilized in active conservation planning. The core challenge addressed is isolating inter-species and intra-species morphological variations associated with injury using a single deep vision model under real-world class imbalances.

Methodology
The study proposes a unified deep learning framework to jointly classify animal species and detect injury-related structural abnormalities from static track images.

• Dataset: A custom dataset of 1,217 images was compiled, merging the Kaggle "Footprint-Animals" dataset (normal tracks) and the Roboflow Universe repository (injured tracks). It covers six species: Bear, Raccoon, Beaver, Lion, Elephant, and Coyote. The data exhibits significant class imbalance, with 1,014 normal samples and 203 injured samples. Images were resized to 224×224 pixels and standardized using ImageNet mean and standard deviation vectors.
• Task Formulation: The primary task is framed as a 12-class fine-grained classification problem, where each label represents a unique combination of species identity and health status (e.g., "Bear-Normal," "Bear-Injured"). A secondary, simplified binary task (Normal vs. Injured) was also evaluated to decouple taxonomic features from injury detection.
• Architectures: Five Convolutional Neural Network (CNN) architectures were benchmarked: ResNet-50, EfficientNet-B0, EfficientNet-B3, MobileNet-V3, and ConvNeXt-V1 Tiny. Models were initialized with ImageNet weights and fine-tuned using PyTorch.
• Optimization: Training utilized the AdamW optimizer with a OneCycle learning rate policy. To address class skew, baseline models used weighted cross-entropy loss. A secondary experiment employed Focal Loss to suppress gradients from majority classes. Data augmentation (flips, rotation, brightness/contrast) was applied strictly to the training split.
• Interpretability: Gradient-weighted Class Activation Mapping (Grad-CAM) was used to verify that models focused on relevant morphological features (e.g., pad clusters, toe spacing) rather than background noise.

Key Results

• Multi-Class Performance: ConvNeXt-V1 Tiny achieved the highest performance in the 12-class setting, reaching an overall accuracy of 78.91% and a macro F1-score of 68%. It demonstrated superior sensitivity to minority classes (e.g., 75% recall for injured bears) compared to other models.
• Model Comparison: EfficientNet-B0 achieved 75.00% accuracy but suffered from a low macro F1-score (55%) due to a systemic bias toward majority "normal" classes, resulting in 0% recall for specific injured cohorts. EfficientNet-B3 underperformed (63.28% accuracy) due to over-parameterization and training instability.
• Loss Function Impact: Implementing Focal Loss improved early convergence for minority classes but slightly reduced global test accuracy to 71.88%, suggesting that while it balances gradients, the dataset size and morphological similarities limit global generalizability.
• Binary Classification: Reformulating the task as a binary classification (Normal vs. Injured) yielded a peak accuracy of 96.88% using ConvNeXt-V1. This model achieved 100% accuracy on normal baselines and sustained >91.3% accuracy on injured configurations.
• Interpretability: Grad-CAM analysis confirmed that top-performing models (ConvNeXt-V1) localized attention precisely on diagnostic anatomical features (pad clusters, claw impressions), whereas lower-performing models exhibited diffuse activation on background substrates.

Significance and Claims
The paper claims to validate a unified deep learning paradigm capable of non-invasive wildlife monitoring without animal handling. Its primary significance lies in demonstrating that:

1. Modern Architectures Matter: Macro-kernel architectures like ConvNeXt are better suited than traditional or lightweight models for capturing the broad spatial context required to differentiate subtle intra-class footprint variations.
2. Task Complexity Trade-off: While fine-grained 12-class models provide comprehensive diagnostic details, binary abstraction models offer significantly higher reliability (96.88% accuracy). The authors suggest that decoupling injury detection from species discrimination removes the bottleneck of overlapping feature representations.
3. Practical Deployment: The study proposes that binary models offer a more practical path for automated field deployment, such as alert systems at watering holes or camera trap arrays, where identifying the presence of injury is the priority over specific species identification.

The authors conclude that while class imbalances and substrate variability remain challenges, this approach minimizes traditional tracking subjectivity and provides a scalable, automated alternative for ecological monitoring. Future work is proposed to expand taxonomic categories and integrate multi-frame temporal inputs to resolve full gait trajectories.