Development of New Aerial Image Datasets and Deep Learning Methods for Waterfowl Detection and Classification

Abstract

Monitoring waterfowl populations and distribution is important for conservation. This paper presents our recent work on creating new aerial image datasets collected by drones and applying and evaluating state-of-the-art deep learning models for waterfowl detection and classification. We collected thousands of aerial images from 10 conservation areas in Missouri, labeled around 600 images with close to 300,000 bird labels, and created 9 datasets with different properties for training and evaluating deep neural network models. Among the models, YOLOv5 performed the best, outperforming Faster R-CNN and RetinaNet. To reduce the amount of labeled data needed for model training, we applied Soft Teacher, a semi-supervised learning method, and obtained slightly better detection performance than supervised learning methods, with just half of the labeled training examples. We trained generic detection models using all datasets containing diverse images and obtained accurate detection results in most cases. For waterfowl classification, we created a dataset of images containing individual waterfowl by cropping them from raw aerial images. We applied several deep learning models to the dataset and obtained promising results.