A digital framework for automated non-invasive waterfowl detection in Carinthia based on high resolution UAS imagery and machine learning

Abstract

Automated waterfowl detection from uncrewed aerial system (UAS; “drones”) imagery has become an important task for various environmental applications such as wildlife monitoring, nature conservation, and habitat mapping. This paper presents a digital framework for automated waterfowl detection using high-resolution UAS imagery and artificial intelligence/machine learning (ML). Several UAS missions in Brenndorf, Carinthia, Austria, were conducted simultaneously with a traditional ground-based waterfowl field survey by an experienced expert. Several data pre-processing steps were applied to optimize digital image data pipelines for the generation of high-quality ML training data. The You Only Look Once (YOLO) open-source computer vision and ML object detection model was used to detect waterfowl in the UAS imagery. A transfer learning approach from a large waterfowl study at the University of New Mexico in collaboration with the U.S. Fish and Wildlife Service was used to further improve the model’s performance. Validation results showed promising performance with 80% and 83% classification accuracy on the waterfowl classes ‘duck’ and ‘swan’, respectively. Finally, a spatial projection model and a visualization approach for the ML-based detection and classification results on a map were implemented. The proposed digital framework for automated waterfowl detection provides promising results for standardization and a new paradigm for waterfowl counting to support and extend traditional wildlife monitoring.