A Comparison Of Machine Learning Methods For Detecting Right Whales From Autonomous Surface Vehicles

This work compares a range of machine learning methods applied to the problem of detecting right whales from autonomous surface vehicles (ASV). Maximising detection accuracy is vital as is minimising processing requirements given the limitations of an ASV. This leads to an examination of the trade-off between accuracy and processing requirements. Three broad types of machine learning methods are explored - convolution neural network (CNNs), time-domain methods and feature-based methods. CNNs are found to give best performance in terms of both detection accuracy and processing requirements. These were also tolerant to downsampling down to 1kHz which gave a slight improvement in accuracy as well as a significant reduction in processing time. This we attribute to the bandwidth of right whale calls which is around 250Hz and so downsampling is able to capture the sounds fully as well as removing unwanted noisy spectral regions.