Table 1.
Machine learning (ML) algorithms used in spectroscopic studies for wildlife science applications.
Fig 1.
A) Raw and B) transformed NIR absorbance spectra (700–2500 nm) were averaged and categorized according to their species designation. The averaged spectra for each species are colored by taxonomic groupings: frogs = green, toads = blue, and salamanders = red. Distinct patterns can be observed between the eleven species investigated across several regions of the NIR spectrum.
Fig 2.
Principal component analysis scores plots for the transformed NIR spectra (700–2500 nm).
A) PCA scores plot including all 11 species indicates distinct patterns between taxonomic orders, with greater spectral differences observed among species belonging to different orders than within orders. Inset: PCA loadings (700–2500 nm) highlighting the dominant peaks that explain the trends present in the scores plot: PC-1 (59%) and PC-2 (33%) explained 92% of database variance. B) PCA scores plot including only the seven anuran species. PC-1 (51%) and PC-2 (31%) explained 82% of the database variance. C) PCA scores plot including only the four caudate species indicate that greater biochemical differences occur between genera than within. PC-1 (79%) and PC-2 (14%) explained 93% of database variance.
Fig 3.
Canonical discriminant plots indicate biochemical separation between individuals belonging to different orders of amphibians (caudates located in the left oval and anurans located in the right oval).
Discriminant plots visualizing biochemical relationships among the four caudates (bottom left) and seven anurans (bottom right) were also investigated. Illustrations by DMC.
Fig 4.
Comparison of mean classification accuracies between the seven models investigated.
Models that differed significantly with regard to classification accuracy (p < 0.05; Tukey HSD) are assigned different letters. KNN = K-nearest neighbors, RF = random forest, PLS = partial least squares, LDA = linear discriminant analysis, GLM = generalized linear model with elastic net regularization, XGB = extreme gradient boosting, and SVM = support vector machine. Minima and maxima are indicated by whiskers, lower and upper interquartile ranges by boxes, medians by black horizontal lines, and means by gold diamonds.
Table 2.
Confusion matrix of results for the support vector machine (i.e., the top-performing model) applied to the combined (anuran + caudate) dataset.
Bolded values indicate the number of accurately classified spectra and non-bolded values indicate the misclassified spectra in the test dataset, respectively. Dotted lines separate major taxonomic orders (Anura vs. Caudata) and reveal that most spectra misclassifications occur within more closely related species.
Fig 5.
Examples of NIR spectra acquisition from live A) tiger salamander (Ambystoma tigrinum) and B) Fowler’s toad (Anaxyrus fowleri). Spectra were non-invasively collected superior to the cloaca on the ventral side for all 11 of the amphibian species investigated. Photos taken by DMC.