Date of Award

December 2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Biomedical Engineering

First Advisor

Hossein Kashani Zadeh

Abstract

When fish species are mislabeled and the fish’s freshness is inaccurately represented, health, environmental, and economic problems can arise for both consumers and the fish industry. Fish are some of the easier foods to mislabel, as the processing of whole fish to fillets makes the visual identification of the species and freshness assessment more challenging. Due to the shortcomings of the current industry standards, confirmed by our interviews with over 20 professionals in all fish supply chain nodes, developing a system that can provide accurate and rapid on-site fraud detection is of necessity. In an attempt to meet these needs, our industry partner has developed a handheld multimodal point spectroscopic (MMPS) device. In this research, we validated that this device can accurately and rapidly assess both the species and freshness of fish fillets when utilizing different machine learning methods to fuse and analyze signals. Additionally, we RGB images taken through smart phones as a more accessible alternative solution. The MMPS device pairs fluorescence (with excitation at 365 and 405 nm) and reflectance (visible near infra-red ~400-1000 nm and short wave near infra-red ~900-1800 nm) into one handheld device. Experiments to validate the potential of this device with different machine learning algorithms, such as neural networks, stacking ensemble methods, linear discriminant analysis, and a novel machine learning framework specialized for many classes have been carried out with hyperspectral line-scan imaging systems, with their data processed to mimic point spectroscopy, as is utilized by our handheld device. These acquired datasets include a robust 40+ species dataset to test fusion data of different spectroscopic modes while classifying species, as well as smaller four sample datasets measured over the course of two weeks, with day number as the ground truth, to validate the ability to classify freshness. Results of over 92% have been achieved for both applications. As the first step to validate the multi-mode point spectroscopy (MMPS) system, our university collaborator collected data using the multi-mode point spectroscopy system from 68 samples and 11 species in both frozen and thawed states, with DNA ground truth for species identification. The machine learning models, with the implementation of our novel dispute model framework, demonstrated the ability to correctly classify the species of the fish over 90% of the time for fish that were frozen, and well as fish that were thawed. 85% was obtained with a dataset that included fish in both states. The final experiment utilizing the MMPS system identifies both the species and freshness of four species and twelve total samples, using nucleotide assays as a method to biologically measure a spoilage indicator, hypoxanthine, of each fish over the course of five days. Spectra were taken from each sample with each of the four modes (30 measurements each day per sample per mode), allowing for 150 spectra to be associated with each species label and freshness label. The species were able to be selected correctly 100% of the time, with the freshness models for each species ranging between 85-99% accuracy when identifying the freshness grade. Smartphones were utilized to take RGB images from two different datasets, one that included three species, and another that comprised of 42 species. Polyethylene wrap and different lighting conditions were also tested to ensure that the images could classify the correct species regardless of the location, and offer a solution that the average consume could use anywhere. The experiment with only three species was able to correctly identify the species 100% of the time, with the 42 species garnering 85% accuracy. The novelty and contribution of this work is to fuse multiple modes of spectroscopy that have been integrated into one hand-held device in a way that allows accurate, rapid, on-site, non-destructive, low-cost, and low skill means of determining species and assessing freshness of fish fillets.

Share

COinS