學生姓名:
廖玉芊
指導教授:
方銘志
學期:
109下
摘 要:
Two methods, fatty acid profiling and phospholipid composition, were applied to the discriminations of wild/farmed and freshwater/seawater cultured salmon. The fatty acid profiles were obtained by GC-FID analysis of lipids from salmons collected from different origins and production methods. The data was processed through various statistic methods such as principle components analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), and other seven machine learning classifiers, namely k16 nearest neighbors (kNN), decisio n tree, suppOrt vector machine (SVM), random forest, artificial neural networks (ANN), naïve Bayes, and AdaBoost. The results showed PCA was able to discriminate wild and farmed salmons, but was not powerful enough to distinguish cultured salmon from different geographical origins such as Canada and Chile. The data was further processed by t-SNE and other seven statistic methods individually.
The best classifier obtained was kNN method by applying only three most important features (16:0, 18:2n6c and 20:3n3+20:4n6) that was able to significantly and clearly separate salmons form different origins. The discrimination between freshwater-cultured and seawater-cultured salmon was accomplished by the analysis of 19 phospholipid fatty acids (PLFAs) in the combination of stepwise discriminate analysis (SDA), PCA, and canonical discriminant analysis (CDA). The results found that four and two PLFAs were significantly related to the cultured salmon caught in different harvest seasons and body size, respectively. Besides, discrimination models through linear discrimination analysis (LDA) and random forest (RF) were created. LDA method achieved an overall correct classification rate of 95.89%, and a predictive rate of 100%. RF method presented equal performance.
01摘要.pdf