Through space and time

Abstract

Modern fishing vessels have a wide range of instruments and sensors on board that are used for active fishing operations, with sonar equipment and echo sounders being among the most common. Sonar allows horizontal observation of the water column, while echo sounders provide more precise underwater environment monitoring. These instruments are useful as they are used today but require a lot of user experience for effective use. Estimating biomass den- sity, fish size, and species is highly demanding, and the existing systems have significant uncertainties. In this thesis, we propose a novel approach to hydroacoustic data analysis that capitalizes on catch reports as annotations for hydroacoustic transects. Com- bining catch messages with the positional attribute of echo data allows us to obtain annotated echo examples that describe the biota within a given loca- tion. The thesis leverages EchoBERT, a BERT-inspired model, as the underlying architecture. To assess the capabilities of the annotations, we evaluate the model using dif- ferent types of models. Both classification and regression tasks are employed, wherein the classification task aims to predict the presence of a species based on catch messages. In contrast, the regression tasks attempt to fit the model to the catch data and generate a distribution of the species. Furthermore, we assess the model considering timestamps. Since the catch messages may not necessarily correspond to the same date as the echo data, we incorporate weighted loss functions that account for the temporal proximity. This approach allows for a closer association during the training process, where the outcome is weighted more heavily for temporally closer labels. Our results provide insight into the characteristics of catch reports as anno- tations, shedding light on their usefulness and limitations. We also uncover potential bias present in the labelled data, where a seasonal fishing activity can be uncovered in the dataset. We also experiment and find the magnitude of difference in collation criterion when finding catch data based on the haversines formula.