Deep learning applied to fish otolith images

• Main Author: Martinsen, Iver
• Format: Master Thesis
• Language: English
• Published: UiT Norges arktiske universitet 2021
• Subjects: VDP::Matematikk og Naturvitenskap: 400::Matematikk: 410::Statistikk: 412; VDP::Mathematics and natural science: 400::Mathematics: 410::Statistics: 412; VDP::Matematikk og Naturvitenskap: 400::Matematikk: 410::Analyse: 411; VDP::Mathematics and natural science: 400::Mathematics: 410::Analysis: 411; Deep Learning; Regression; Classification; Statistical methods; STA-3900; Arctic; Greenland; Arctic cod; Northeast Arctic cod
• Online Access: https://hdl.handle.net/10037/25212

This thesis is concerned with classification and regression using deep learning applied to fish otolith images. Otoliths (earstones) are calcified structures in the inner ear of vertebrates, and are used, for instance, in fish stock assessment and fish age determination. We use convolutional neural networks – a class of deep learning models - on two specific problems: discrimination between Northeast Arctic Cod and Norwegian Coastal Cod, and age determination of Greenland halibut. In relation to classification and regression, we are also concerned with the usage of cross-validation procedures such as k*l-fold cross-validation, to obtain reliable test results. We obtain test results for all available data, and we argue for the usage of cross-validation on the bases of variations in test results. Furthermore, feature relevance attribution methods are discussed and compared, which aims at explaining outputs from deep learning models by attributing relevance scores to the input. These comparisons are conducted using image input heatmaps produced by methods such as gradient saliency maps, guided backpropagation, and integrated gradients, along with two proposed variations of those techniques.