| Summary: | Knowledge of fat in salmon is extremely important to salmon breeder and the whole salmon industry. By monitoring fat in salmon fillet, huge amount of money will be saved. Several methods are available to determine fat in salmon fillets. Stofnfiskur Iceland decided to buy the NIR instrument Qmonitor which was installed in there slaughter line. When applying existing prediction model to results obtained by Qmonitor the prediction of fat was wrong. Aim of this thesis is to develop a new valid prediction model which will be applied to results obtained by the NIR instrument Qmonitor when measuring fish from all families in the nucleus of Stofnfiskur for breeding purposes. This thesis will provide background of NIR, breeding and problems of modeling fat in salmon fillet. Main goal is to discuss methods needed to explore the data, develop prediction model and validate the prediction model obtained. Use of recently developed CPLS will then be introduced in order to reduce the prediction error of existing methodology when creating prediction model. All methods will be compared and there qualities and drawback discussed. Three datasets are presented in the thesis were two of them where made for this thesis and one comes from paper defining methods used when modeling QMonitor data. In the paper where the method of picking out five $14$ mm plugs from the fillet to capture the variation of fat in the fillet a RMSEP value reported was $1.96$. By using Canonical Partial Least Squares with the additional response a location of the plug, the RMSEP of the same dataset was $1.75$. On the dataset made for this thesis to develope prediction model for the QMonitor in Iceland CPLS had the best performance obtaining RMSEP value of $1.8$. Additional values which improved the prediction model where additional information about the plugs such as thickness of the plug, moisture in the plug and weight of the plug. Stofnfiskur
|
|---|