Drone Monitoring and Behavioral Analysis of White-Beaked Dolphins (Lagenorhynchus albirostris)

Abstract

Highlights: What are the main findings? Drone imagery combined with AI successfully identified key behavioral types of white-beaked dolphins (Traveling, Milling, Respiration), with high precision for Traveling but limited accuracy for Milling and especially Respiration. Use of RGB (color) data improved classification performance compared to grayscale, although visual disturbances such as birds, waves, and reflections frequently caused misclassifications. What is the implication of the main finding? Findings demonstrate that drones coupled with machine learning offer a non-invasive tool for monitoring marine mammal behavior, but methodological refinements are needed for complex or short-duration behaviors. The approach has strong potential to support more efficient and ethically sound marine mammal management, particularly through multimodal AI systems and larger, balanced datasets in future studies. Marine mammals serve as indicator species for environmental and human health. However, they are increasingly exposed to pressure from human activities and climate change. The white-beaked dolphin (Lagenorhynchus albirostris) (WBD) is among the species negatively affected by these conditions. To support conservation and management efforts, a deeper understanding of their behavior and movement patterns is essential. One approach is drone-based monitoring combined with artificial intelligence (AI), allowing efficient data collection and large-scale analysis. This study aims to: (1) investigate the use of drone imagery and AI to monitor and analyze marine mammal behavior, and (2) test the application of machine learning (ML) to identify behavioral patterns. Data were collected in Skjálfandi Bay, Iceland, between 2021 and 2023. Three behavioral types were identified: Traveling, Milling, and Respiration. The AI_RGB model showed high performance on Traveling behavior (precision 92.3%, recall 96.9%), while the AI_gray model achieved higher precision (97.3%) but much lower recall (9.5%). The model struggled to classify Respiration accurately (recall 1%, F1-score 2%). A key challenge was misidentification of WBDs due to visual overlap with birds, waves, and reflections, resulting in high false positive rates. Multimodal AI systems may help reduce such errors in future research.