Drones and convolutional neural networks facilitate automated and accurate cetacean species identification and photogrammetry

Abstract

The flourishing application of drones within marine science provides more opportunity to conduct photogrammetric studies on large and varied populations of many different species. While these new platforms are increasing the size and availability of imagery datasets, established photogrammetry methods require considerable manual input, allowing individual bias in techniques to influence measurements, increasing error and magnifying the time required to apply these techniques. Here, we introduce the next generation of photogrammetry methods utilizing a convolutional neural network to demonstrate the potential of a deep learning-based photogrammetry system for automatic species identification and measurement. We then present the same data analysed using conventional techniques to validate our automatic methods. Our results compare favorably across both techniques, correctly predicting whale species with 98% accuracy (57/58) for humpback whales, minke whales, and blue whales. Ninety percent of automated length measurements were within 5% of manual measurements, providing sufficient resolution to inform morphometric studies and establish size classes of whales automatically. The results of this study indicate that deep learning techniques applied to survey programs that collect large archives of imagery may help researchers and managers move quickly past analytical bottlenecks and provide more time for abundance estimation, distributional research, and ecological assessments.