Beyond Accuracy: The ML Field Planner's Framework for AI Model Selection in Conservation

Artificial intelligence (AI) and machine learning (ML) are increasingly applied across various scientific domains, supported by high-quality datasets, domain-specific knowledge, modern cyberinfrastructure (spanning from edge to core devices), and AI expertise. Deploying these tools—whether at the edge for rapid inference or at the center for large-scale training—requires an understanding of model architectures, resource needs, and infrastructure constraints; however, these demands can divert time from ecological research.

When the data being generated are images, ecologists often turn to state-of-the-art computer vision models, as object detection and classification are well-developed research areas. Many domains can leverage such models with minimal customization. The YOLO can be fine-tuned for tasks in animal ecology, digital agriculture, or smart mobility. The paper "Perspectives in Machine Learning for Wildlife Conservation " [1] demonstrates how MegaDetector addresses bottlenecks by filtering out blank images, reducing human review to under 15%, and saving storage. Yet, selecting and testing models—balancing accuracy, resources, and deployment—remains a tedious but necessary task.

In the ML Field Planner [2], we introduced a framework to help ecologists to use benchmark datasets in selecting models based on accuracy, latency, energy use, and memory footprint, and receive performance-driven recommendations. This poster builds on that work by outlining the key components of the ML Field Planner, the range of evaluation metrics it incorporates, and example use cases demonstrating how these metrics can guide deployment choices. We also highlight scenarios where ecologists may deploy different models for different purposes—for example, using YOLO to filter out blank images or to detect a target species and trigger an autonomous drone for follow-up observations.

Model selection is informed not only by accuracy but also by additional performance indicators such as false positives, precision, recall, and mean Average Precision (mAP). For example:
• YOLOv11 Nano – Precision 74.56%, Recall 84%, False Positives 11, Size 1.5 M parameters / 5.53 MB: effective as a lightweight blank-image filter at the edge.
• YOLOv5 Extra Large v5 – Precision 91.19%, Recall 68%, False Positives 14, Size 97 M parameters / 311.72 MB: suited for high-precision species classification on near-edge devices.

By integrating these metrics, the ML Field Planner supports testing workflows such as:
• deleting blanks and identifying species;
• deleting blanks and sending non-blank images for processing; and
• detecting a target species and deploying a drone for targeted imagery.
Evaluated against performance metrics and device limits, these workflows enable ecologists to make efficient, informed deployment decisions that align AI use with conservation goals while minimizing technical overhead.