There are multiple research groups who have built species-classification models for fish based upon infra-red (IR) spectra. Most models, ultimately, are driven by differences in the fingerprints of IR absorbing compounds such as lipids and amides.
This study (purchase required) highlights the importance of standardising the part(s) of the fish (tail, belly etc) selected to train the model, and the different outcomes that can be obtained by using models that are trained on different parts of the fish.
The researchers examined whether intra-fillet anatomical heterogeneity influenced the ability of PLS-DA models based on ATR-FTIR spectra to discriminate three pooled high-value salmon species (Salmo salar, Oncorhynchus kisutch, and Oncorhynchus tshawytscha) from rainbow trout (Oncorhynchus mykiss). A total of 122 spectra from the back, belly, and tail regions of 14 fish were evaluated using outer Leave-One-Fish-Out validation with inner fish-grouped latent-variable selection. The final full-data models used 3, 3, and 4 latent variables for back, belly, and tail, respectively; model size was selected separately within each outer validation fold. Fish-level balanced error rates were 0, 0.063, and 0.250, and AUC-ROC values (based on relatively small cohorts) were 1.000, 0.979, and 0.917, respectively.
They found that the tail model showed lower specificity and less stable latent-variable selection under fish-level internal validation. VIP analysis identified 366 wavenumbers shared across all three regional sets, representing 12.99% of their union, while most selected variables remained region-dependent. Predictive performance differed across anatomical sampling locations.
They conclude that, for their application, dorsal and ventral muscle are less suited as sampling sites for routine deployment.
Photo by Creab ThePolymath on Unsplash
Comments