A general deep learning model for bird detection in high-resolution airborne imagery.

Area of Science:

• Ecology
• Artificial Intelligence
• Computer Vision

Background:

• Ecological studies rely on understanding individual distribution and behavior.
• Computer vision, particularly deep neural networks, can detect objects in imagery.
• Supervised models for ecological monitoring face challenges like large data requirements, technical expertise, and overfitting.

Purpose of the Study:

• To develop a generalized artificial intelligence model for bird detection in ecological monitoring.
• To overcome limitations of traditional supervised models in terms of data, expertise, and computational resources.
• To enable large-scale, automated detection of individual organisms across diverse species and ecosystems.

Main Methods:

• Developed a general bird detection model using over 250,000 annotations from 13 global projects.
• Applied the model to novel aerial data without local training.
• Fine-tuned the general model with a small set of local annotations (1000).

Main Results:

• The general model achieved over 65% recall and 50% precision on novel aerial data without local training.
• Fine-tuning with 1000 local annotations improved performance to an average of 84% recall and 69% precision.
• Retraining from the general model enhanced local predictions, reduced training time, and increased stability.

Conclusions:

• General models for detecting broad classes of organisms using airborne imagery are feasible.
• These models can significantly reduce the effort, expertise, and computational resources for ecological monitoring.
• The developed approach has the potential to transform the scale of ecological data collection and research questions.