Machine Learning-Based Non-Destructive Classification of Tilapia Flesh Using Spectroscopic Techniques
Journal
2025 Asabe Annual International Meeting
Date Issued
2025
Author(s)
Abstract
Tilapia, the leading aquaculture species in Taiwan, faces quality challenges due to internal flesh abnormalities, such as those caused by Streptococcosis, which are difficult to detect until it is sectioned. This study addresses these challenges by applying hyperspectral imaging (HSI) combined with machine learning techniques for meat quality assessment, thereby reducing disputes between aquaculture farmers and fillet processors. Sectioned fillet samples were provided by several fishery production cooperatives in Taiwan, including 8 normal fillets and 13 fillets affected by Streptococcosis. Comprehensive spectral data were captured from both the inner flesh and outer surfaces of the fillets using HSI. Subsequently, down-sampling was performed on the HSI dataset to reduce data complexity and enable rapid classification, yielding a total of 277,438 spectra. Four machine learning methods were used to classify fillets, with model performance compared across datasets. Key findings demonstrate that for HSI, one-dimensional convolutional neural network (1D-CNN) achieved superior accuracies of 0.97 for the outer surface and 0.99 for the inner flesh, effectively capturing subtle features affected by Streptococcosis. Additionally, a brief comparative analysis was conducted using Near-Infrared (NIR) spectroscopy, which showed lower accuracy but offered advantages in rapid data acquisition and potential portability for on-site assessments. The findings highlight the superior capability of HSI for precise, spatially-resolved analysis, underscoring its potential to significantly enhance quality assessment processes in the aquaculture industry. This innovative approach addresses critical challenges in tilapia quality assessment, offering a pathway for improved industry practices.
Event(s)
2025 American Society of Agricultural and Biological Engineers Annual International Meeting, ASABE 2025
Subjects
hyperspectral imaging
machine learning
Non-destructive assessment
Publisher
American Society of Agricultural and Biological Engineers
Description
2025 American Society of Agricultural and Biological Engineers Annual International Meeting, ASABE 2025, Toronto, Ontario, Canada July 13-16, 2025
Type
conference paper