On-Land Pinniped Classification of Multiple Species and Demographic Classes on Multiple Substrates Using Deep Learning and Aerial Imagery

Abstract

Pinniped abundance and demographic monitoring is the foundation of informed decisions about conservation, management and protection. However, current aerial and on the ground monitoring techniques are generally resource intensive, may suffer detection error and can present hazards to on-site surveyors. Advances in deep learning in combination with high quality aerial imagery can minimise the safety risks and resources required by current monitoring techniques and allow for the quick analysis of legacy and contemporary images for pinniped species on various substrates. We used aerial images (N = 218) collected from the California Channel Islands to train a Retinanet50 model to detect elephant seals hauled out on the sandy beach and label them as either ‘bull’, ‘cow’ or ‘pup’. Using the elephant seal model as a starting point, we fine-tuned this model to detect harbour seals on a variety of substrates using a limited number of images (N = 13). Both models achieved high accuracy with mean average precisions of 94% and 95% respectively. The process of fine tuning for a second species on different substrates was significantly faster than the creation of the initial model, reducing both model training and data labelling costs. This approach is automatable and would increase accuracy, improve timeliness, decrease the resources required to monitor pinniped populations at the age class level on variable substrates, increase count accuracy and improve human safety in rugged terrain.