Limits...
Multispectral imaging for determination of astaxanthin concentration in salmonids.

Dissing BS, Nielsen ME, Ersbøll BK, Frosch S - PLoS ONE (2011)

Bottom Line: A PLSR model was calibrated to predict astaxanthin concentration from novel images, and showed good results with a RMSEP of 0.27.For comparison a similar model were built for normal color images, which yielded a RMSEP of 0.45.The acquisition speed of the multispectral imaging system and the accuracy of the PLSR model obtained suggest this method as a promising technique for rapid in-line estimation of astaxanthin concentration in rainbow trout fillets.

View Article: PubMed Central - PubMed

Affiliation: Department of Informatics and Mathematical Modeling, Technical University of Denmark, Kgs. Lyngby, Denmark.

ABSTRACT
Multispectral imaging has been evaluated for characterization of the concentration of a specific cartenoid pigment; astaxanthin. 59 fillets of rainbow trout, Oncorhynchus mykiss, were filleted and imaged using a rapid multispectral imaging device for quantitative analysis. The multispectral imaging device captures reflection properties in 19 distinct wavelength bands, prior to determination of the true concentration of astaxanthin. The samples ranged from 0.20 to 4.34 g per g fish. A PLSR model was calibrated to predict astaxanthin concentration from novel images, and showed good results with a RMSEP of 0.27. For comparison a similar model were built for normal color images, which yielded a RMSEP of 0.45. The acquisition speed of the multispectral imaging system and the accuracy of the PLSR model obtained suggest this method as a promising technique for rapid in-line estimation of astaxanthin concentration in rainbow trout fillets.

Show MeSH
Directions of maximum variation in the autoscaled meanspectra.A scoreplot for a principal component analysis of the 19 dimensional mean spectra shows a definite trend in the the data along the first principal component which accounts for 80% of the variation in the dataset. The second principal component on the y axis accounts for a total of 14%. Each label in the plot represents a multispectral image of a trout fillet, and the number is the concentration in the corresponding fillet in g.
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC3091859&req=5

pone-0019032-g009: Directions of maximum variation in the autoscaled meanspectra.A scoreplot for a principal component analysis of the 19 dimensional mean spectra shows a definite trend in the the data along the first principal component which accounts for 80% of the variation in the dataset. The second principal component on the y axis accounts for a total of 14%. Each label in the plot represents a multispectral image of a trout fillet, and the number is the concentration in the corresponding fillet in g.

Mentions: Figure 9 shows a score plot (PC1 versus PC2) of a PCA of the entire dataset after autoscaling. The plot shows a clear trend in the first component describing 80% of the variation with few outliers. Based on outlier diagnostics seven samples were categorized as outliers and removed from the data set prior to further analysis. All outliers were characterized by bad filleting. Table 2 shows the results of the final PLSR model for astaxanthin prediction based on a multispectral image. The reported RMSEP of 0.26 from a 7 component PLSR model is based on an independent test set while the model itself was cross validated on a trainingset. The cross validation showed a minimum generalization error when using 7 PLS components (Figure 10), which together with a total variance description on the response variable of 91% led to the choice of 7 components in the model. The variance decription percentages for the 7 components in the response variable were 48, 69, 75, 83, 86, 89 and 91% which shows that the performance of the model drastically increases with the first 4 components. The loadings for these components are shown in Figure 11. The first 2 components clearly show high response in the area around the absorption peak of astaxanthin in the blue/green area of the visual spectrum. The components naturally reflect the areas of largest variation in the preprocessed spectra shown in Figure 8. Table 2 also contain a PLSR model fitted on the same data after transformation to sRGB images. This means only a total of 3 variables exists for the regression problem which could basically be handled using a full multiple linear regression. This was tested together with a PLS model, which was found to yield equal results. The RGB model in Table 2 is seen to have a higher RMSEP value, indicating reduced prediction abilities than the multispectral mode. Among the two models the multispectral models has best variance description with an value of 0.86 versus and value of 0.66 for the RGB model. Furthermore the variance in the residuals is seen to be smaller for the multispectral model with a standard error of 0.02 against 0.05 for the RGB model. Previously similar techniques has been investigated using VIS spectroscopy and digital photography in [36], for calibration against chemically measured astaxathin. VIS spectroscopy was here found to have a correlation coefficient of 0.92 with a RMSEP of 0.42. For digital photography a correlation of 0.92 was found together with an RMSEP of 0.41. They reported a cross validation error based on all samples, meaning this error was used to chose the correct number of components. Compared to our results, we managed to achieve an RMSEP of 0.27 on an independent testset, to get a truly unbiased model, for the multispectral images.


Multispectral imaging for determination of astaxanthin concentration in salmonids.

Dissing BS, Nielsen ME, Ersbøll BK, Frosch S - PLoS ONE (2011)

Directions of maximum variation in the autoscaled meanspectra.A scoreplot for a principal component analysis of the 19 dimensional mean spectra shows a definite trend in the the data along the first principal component which accounts for 80% of the variation in the dataset. The second principal component on the y axis accounts for a total of 14%. Each label in the plot represents a multispectral image of a trout fillet, and the number is the concentration in the corresponding fillet in g.
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC3091859&req=5

pone-0019032-g009: Directions of maximum variation in the autoscaled meanspectra.A scoreplot for a principal component analysis of the 19 dimensional mean spectra shows a definite trend in the the data along the first principal component which accounts for 80% of the variation in the dataset. The second principal component on the y axis accounts for a total of 14%. Each label in the plot represents a multispectral image of a trout fillet, and the number is the concentration in the corresponding fillet in g.
Mentions: Figure 9 shows a score plot (PC1 versus PC2) of a PCA of the entire dataset after autoscaling. The plot shows a clear trend in the first component describing 80% of the variation with few outliers. Based on outlier diagnostics seven samples were categorized as outliers and removed from the data set prior to further analysis. All outliers were characterized by bad filleting. Table 2 shows the results of the final PLSR model for astaxanthin prediction based on a multispectral image. The reported RMSEP of 0.26 from a 7 component PLSR model is based on an independent test set while the model itself was cross validated on a trainingset. The cross validation showed a minimum generalization error when using 7 PLS components (Figure 10), which together with a total variance description on the response variable of 91% led to the choice of 7 components in the model. The variance decription percentages for the 7 components in the response variable were 48, 69, 75, 83, 86, 89 and 91% which shows that the performance of the model drastically increases with the first 4 components. The loadings for these components are shown in Figure 11. The first 2 components clearly show high response in the area around the absorption peak of astaxanthin in the blue/green area of the visual spectrum. The components naturally reflect the areas of largest variation in the preprocessed spectra shown in Figure 8. Table 2 also contain a PLSR model fitted on the same data after transformation to sRGB images. This means only a total of 3 variables exists for the regression problem which could basically be handled using a full multiple linear regression. This was tested together with a PLS model, which was found to yield equal results. The RGB model in Table 2 is seen to have a higher RMSEP value, indicating reduced prediction abilities than the multispectral mode. Among the two models the multispectral models has best variance description with an value of 0.86 versus and value of 0.66 for the RGB model. Furthermore the variance in the residuals is seen to be smaller for the multispectral model with a standard error of 0.02 against 0.05 for the RGB model. Previously similar techniques has been investigated using VIS spectroscopy and digital photography in [36], for calibration against chemically measured astaxathin. VIS spectroscopy was here found to have a correlation coefficient of 0.92 with a RMSEP of 0.42. For digital photography a correlation of 0.92 was found together with an RMSEP of 0.41. They reported a cross validation error based on all samples, meaning this error was used to chose the correct number of components. Compared to our results, we managed to achieve an RMSEP of 0.27 on an independent testset, to get a truly unbiased model, for the multispectral images.

Bottom Line: A PLSR model was calibrated to predict astaxanthin concentration from novel images, and showed good results with a RMSEP of 0.27.For comparison a similar model were built for normal color images, which yielded a RMSEP of 0.45.The acquisition speed of the multispectral imaging system and the accuracy of the PLSR model obtained suggest this method as a promising technique for rapid in-line estimation of astaxanthin concentration in rainbow trout fillets.

View Article: PubMed Central - PubMed

Affiliation: Department of Informatics and Mathematical Modeling, Technical University of Denmark, Kgs. Lyngby, Denmark.

ABSTRACT
Multispectral imaging has been evaluated for characterization of the concentration of a specific cartenoid pigment; astaxanthin. 59 fillets of rainbow trout, Oncorhynchus mykiss, were filleted and imaged using a rapid multispectral imaging device for quantitative analysis. The multispectral imaging device captures reflection properties in 19 distinct wavelength bands, prior to determination of the true concentration of astaxanthin. The samples ranged from 0.20 to 4.34 g per g fish. A PLSR model was calibrated to predict astaxanthin concentration from novel images, and showed good results with a RMSEP of 0.27. For comparison a similar model were built for normal color images, which yielded a RMSEP of 0.45. The acquisition speed of the multispectral imaging system and the accuracy of the PLSR model obtained suggest this method as a promising technique for rapid in-line estimation of astaxanthin concentration in rainbow trout fillets.

Show MeSH