Noninvasive Low‐cost Method to Identify Armadillos' Burrows: A Machine Learning Approach

Abstract

Having accurate information about population parameters of armadillos (Mammalia, Cingulata) is essential for the conservation and management of the taxon, most species of which remain poorly studied. We investigated whether we could accurately identify 4 armadillo species (Euphractus sexcinctus, Dasypus novemcinctus, Cabassous tatouay, and Cabassous unicinctus) based on burrow morphometry. We first selected published studies that reported measurements of width, height, and angle of the burrows used by the 4 species of armadillos. Then, using such data we simulated burrow measurements for each of the 4 species of armadillos and we created predictive models through supervised machine learning that were capable of correctly identifying the species of armadillos based on their burrows' morphometry. By using classification algorithms such as Random Forest, K-Nearest Neighbor, Support Vector Machine, Naive Bayes, and Decision Tree C5.0, we achieved the overall accuracy for the classification task by about 71%, including an overall Kappa index by about 61%. Euphractus sexcinctus was the most difficult species to discriminate and classify (approximately 68% of accuracy), whereas C. unicinctus was the easiest to discriminate (approximately 93% of accuracy). We found that it was possible to identify similar-sized armadillos based on the measurements of their burrows described in the literature. Finally, we developed an R function (armadilloID) that automatically identified the 4 species of armadillos using burrow morphology. As the data we used represented all studies that reported the morphometry of burrows for the 4 species of armadillos, we can generalize that our function can predict armadillo species beyond our data.