Erratum: Deep learning improves taphonomic resolution: High accuracy in differentiating tooth marks made by lions and jaguars (Journal of the Royal Society Interface (2020) 17 (20200446) DOI: 10.1098/rsif.2020.0446)

Abstract

In our previous paper [1], we presented some convolutional neural network (CNN) models to classify images of tooth scores made by lions and jaguars through deep learning computer vision. In that work, we reached an accuracy of 82% of the testing set correctly classified. However, such an accuracy is biased, since the original sample was highly unbalanced. The accuracy reported was impacted by a better classification of the larger lion sample than the smaller jaguar sample. In our original study, to compensate for the unbalanced toothmark samples, we created a more balanced subsample composed of 42 images of tooth marks from jaguars and 42 tooth marks from lions. Thiswas done by shuffling the original image dataset and randomly sampling 42 images of each agent. In this smaller sample, accuracy decreased slightly. When confusion matrices for this subsample were inspected, it appeared that the classification of tooth marks from both agents was balanced. For the most successful model (VGG19; accuracy = 75.6), the F1-score factor was 0.71. We neglected checking the accuracy balance in the larger sample. We did that posteriorly and realized a divergence between precision and recall in our models. In the larger sample, tooth marks of lionswerewell classified, but a significant portion of those of jaguarswere misclassified. The reason for this is that a fewof the toothmarksmade by lions display microscopic features that are very similar to those documented in jaguar tooth scores. When using a small randomly sampled set of images, the probability of including that minor part of the lion sample is small and, hence, the similar values obtained for accuracy and F1-score factors. However, when using the much larger lions sample, that jaguar-looking portion of tooth scores is enough to produce a low precision/recall for the jaguar testing sample, because the algorithm sees those marks similar to those documented in lions. A subsequent analysis of multiple carnivore tooth marks underscored this problem by showing systematic misclassification of the jaguar tooth scores [2]. They were mostly classified as lion tooth marks. We thought that this indicated that both types of tooth marks were situated in different parts of a general felid tooth mark spectrum, but that they overlapped enough to make their differentiation difficult (against our own previous work). Additionally, we also realized that in most of the transfer knowledge models used in Jiménez-García et al.’s study [1], we used the same pre-processing standard function, instead of using the model-specific pre-processing functions. We thought these might have produced different results. In the multiple carnivore study, we also realized that augmentation, usually considered a panacea for avoiding overfitting and producing higher accuracy models [3], did not universally do so, and several models yielded higher accuracy when not using image augmentation. For these reasons, we considered adequate to address whether the unbalanced classification problem of the published lion-jaguar models, which rendered their utility rather limited, could be overcome by using non-augmented architectures and using the pre-processing functions specifically designed for each of the transfer knowledge models.We also thought that ensemble learning, which is known to produce more balanced results, could also improve the precision–recall imbalance documented in our previous modelling. Here, we present the results, which correct the problems of the previously published models by producing more balanced classifications and also by achieving higher accuracy than previously reported.