LizardNet: A mobile hybrid deep learning tool for classification of 3D representations of Amazonian lizards

Abstract

Image classification is a highly significant field in machine learning (ML), especially when applied to address longstanding and challenging issues in the biological sciences. In this study, we present the development of a hybrid deep learning-based tool suitable for deployment on mobile devices. This tool is aimed at processing and classifying three-dimensional samples of endemic lizard species from the Amazon rainforest. The dataset used in our experiment was collected at the Museu Paraense Emilio Goeldi (MPEG), Belem-PA, Brazil, and comprises three species: a) Anolis fuscoauratus; b) Hoplocercus spinosus; and c) Polychrus marmoratus. We compared the effectiveness of four artificial neural networks (ANN) for feature extraction: a) MobileNet; b) MobileNetV2; c) MobileNetV3Small; and d) MobileNetV3Large. Additionally, we evaluated five classical ML models for classifying the extracted patterns: a) Support Vector Machine (SVM); b) GaussianNB (GNB); c) AdaBoost (ADB); d) K-Nearest Neighbors (KNN); and e) Random Forest (RF). Our most effective model, MobileNetV3-Small + Linear SVM, achieved an accuracy of 0.948 and a f1-score of 0.955. Notably, it not only proved to be the least complex model among all combinations but also demonstrated the best performance after a statistical comparison. These results indicate that the combination of deep learning (DL) models with less complex classical ML algorithms, which have a lower error propensity, emerges as a viable and efficient technique for classifying three-dimensional lizard species samples. Such an approach facilitates taxonomic identification work for professionals in the field and provides a tool adaptable for integration into mobile data recording equipment, such as smartphones. Author summaryThe taxonomic classification of lizards requires an exceptional level of knowledge and attention to minute details beyond the ordinary to accurately categorize specimens. Such tasks impose significant mental and visual costs on humans, unlike computer vision algorithms capable of extracting visual patterns from images imperceptible to the human eye. In this research, we utilized a dataset from the herpetarium of the Emilio Goeldi Museum in Belem-PA, Brazil. The data were self-captured, with each sample comprised of three photos: dorsal, lateral, and ventral views of each specimen. The sample size was constrained by the quality and abundance of preserved specimens, necessitating the application of a data augmentation method on the pre-separated training and validation sets. This augmentation led to a considerable increase in the number of samples per species, from a few dozen to several hundred. Our experimental approach involved utilizing pre-trained neural networks to extract 3D sample characteristics, subsequently classified using classical machine learning algorithms. This hybrid strategy was adopted due to the nature of data collection and synthetic data augmentation. Our method enables specimen identification through three-dimensional representations, allowing for a more comprehensive utilization of morphological information by the model.