Identification of high-frequency periodic acoustic tags with deep learning

Abstract

Marine life researchers use the concept of fish tracking to determine the activity andbehaviour of fish. It allows the researchers to recognize the valuable biological andphysical support systems required by fish species at various life stages. There havebeen several advancements in fish tracking using different approaches that includeelectronic tags and acoustic tags. In acoustic fish tracking, the fish are equipped withacoustic tags that are either externally attached or surgically implanted into the fish.These tags can be tracked by a receiver placed within a range underwater. The tagsemit high-frequency signals omnidirectionally at regular intervals, and these signalsare recorded as pings in the receiver. The tags are identified by the delay betweenthe pulses, which is typically between 1 second and 10 seconds. The tags can also beconfigured to emit two closely spaced pulses providing a new way to identify the tags.The pings recorded by the receiver are represented as acoustic time series datathat is analyzed to determine which tag the ping originated from. This analysis ofidentifying the pings (marking) from high-frequency tags which use a double pulseencoding scheme is currently done at Innovasea with the help of a visual analyticssystem called ‘MarkTags’, where the marking is done automatically within the soft-ware, or the user manually tunes the data to a particular tag period and marks thepings which consume much manual effort. We proposed a machine learning solutionfor identifying the tags using deep learning.Our work discusses a novel approach to finding the pings by segmenting the im-ages created from the pings. The pings are created as images using a 2D-histogramapproach in which the pings are represented as pixels in an image. These imagesare segmented for the pings from the tag using a neural network called UNet. Wedeveloped a model that is trained on data recorded in different conditions. When thetrained model was asked to identify the tags not seen during training, it could do sowith an accuracy of over 95% and is close to human-level annotations. Our exper-imental results show that the machine learning approach matches the human-levelperformance, which can replace human intervention to a great extent.