Intra-African Geographic Domain Shift in Wildlife Camera Trap Species Classification: A Comparative Study of Supervised and Zero-Shot Foundation Models

Technical Summary: Intra-African Geographic Domain Shift in Wildlife Camera Trap Species Classification

Problem Statement
While camera trap networks like Snapshot Safari have generated millions of labeled wildlife images across Africa, facilitating the training of deep learning models for automated species classification, a critical gap remains in understanding model generalizability. Specifically, the performance degradation caused by geographic domain shift—deploying models trained in one African region to another within the continent—is poorly understood. Conservation practitioners often face the challenge of deploying models across diverse ecosystems (e.g., from East to Southern Africa) without the resources to collect new labeled data for every new location.

Methodology
This study presents the first systematic evaluation of intra-African geographic domain shift using three distinct model architectures that interact with the Snapshot Serengeti dataset in fundamentally different ways:

1. Supervised Baseline (BEiTV2): The model is fine-tuned directly on Serengeti images, representing a standard supervised learning approach.
2. Retrieval-Based Approach (DINOv2 + FAISS): This method utilizes Serengeti images as a retrieval index within a FAISS vector database. The model weights are not updated; classification is achieved by retrieving the most similar Serengeti images to the query.
3. True Zero-Shot Foundation Model (BioCLIP): This architecture receives no Serengeti training data whatsoever, relying entirely on its pre-trained foundation capabilities to classify species.

Experimental Design
All three models were evaluated on two distinct Southern African test sets: Snapshot Kgalagadi and Snapshot Kruger, as well as on a set of locally collected wildlife photographs from Botswana. The study encompasses eight specific experiments designed to isolate variables affecting domain shift:

• In-domain baselines.
• Cross-dataset transfer performance.
• Data scaling effects.
• The impact of MegaDetector preprocessing.
• Comparisons between grayscale and color image conditions.
• Per-species transfer analysis to identify which species are most or least affected by domain shift.

Key Contributions and Results
The paper provides the first empirical characterization of intra-African domain shift across both supervised and zero-shot architectures. By comparing BEiTV2, DINOv2, and BioCLIP, the study quantifies how different learning paradigms handle the ecological and visual variations between East and Southern African ecosystems. The results offer a comparative analysis of:

• The efficacy of fine-tuning versus zero-shot inference in a conservation context.
• The utility of retrieval-based methods as a middle ground.
• The robustness of models under varying preprocessing conditions (e.g., MegaDetector) and image modalities (color vs. grayscale).
• Specific species-level vulnerabilities to geographic domain shift.

Significance
The primary significance of this work lies in its provision of practical guidance for conservation AI practitioners. It addresses the real-world constraint of deploying models across the diverse ecosystems of Southern Africa without the necessity of collecting new labeled data for every deployment site. By systematically characterizing the limits and capabilities of current foundation models and supervised baselines in this specific geographic context, the study enables more informed decisions regarding model selection and deployment strategies for wildlife monitoring across the African continent.