Evaluating the Efficacy of Drone-Based Thermal Images for Measuring Wildlife Abundance and Physiology

Abstract

Although drones are a promising alternative to traditional wildlife monitoring methods, validation efforts are needed to quantify the accuracy of abundance and distribution estimates obtained from using drones. We used drones equipped with high-resolution Red-Green-Blue (RGB) and thermal cameras, coupled with machine learning techniques, to assess the abundance and thermal physiology in northern elephant seals ( Mirounga angustirostris ). Aerial images of 3415 seals and measurements of ambient air temperature, wind speed, and time of day were collected during nighttime and daytime drone flights (N = 24). Two-dimensional polygons and surface temperatures of seals were measured from the images. Machine learning algorithms were applied to detect seals in the imagery, and model performance was evaluated. Detection was more accurate using RGB images (machine learning averaged 6.8% lower than human counts) than thermal images (16.6%). However, thermal images were useful for determining that time of day and ambient temperature (but not wind speed or body size) influenced seal external skin temperature. RGB and thermal cameras have different strengths and weaknesses that should be considered when designing research studies. Our study demonstrates that integrating drones, thermal imaging, and machine learning can promote faster, safer, cheaper, less disruptive, and more accurate wildlife monitoring and conservation efforts.