Department of Astronomy, Space Science and Geology, Chungnam National University, Daejeon, Korea.
Department of Geological Sciences, Chungnam National University, Daejeon, Korea.
PLoS One. 2023 Nov 29;18(11):e0293020. doi: 10.1371/journal.pone.0293020. eCollection 2023.
Large ornithopod dinosaur footprints have been confirmed on all continents except Antarctica since the 19th century. However, oversplitting problems in ichnotaxa have historically been observed in these footprints. To address these issues and distinguish between validated ichnotaxa, this study employed convolutional neural network-based Xception transfer learning to automatically classify ornithopod dinosaur tracks. The machine learning model was trained for 162 epochs (i.e., the number of full cycles of all training data through the model) using 274 data images, excluding horizontally flipped images. The trained model accuracy was 96.36%, and the validation accuracy was 92.59%. We demonstrate the performance of the machine learning model using footprint illustrations that are not included in the training dataset. These results show that the machine learning model developed in this study can properly classify footprint illustration data for large ornithopod dinosaurs. However, the quality of footprint illustration data (or images) inherently affects the performance of our machine learning model, which performs better on well-preserved footprints. In addition, because the developed machine-learning model is a typical supervised learning model, it is not possible to introduce a new label or class. Although this study used illustrations rather than photos or 3D data, it is the first application of machine-learning techniques at the academic level for verifying the ichnotaxonic assignments of large ornithopod dinosaur footprints. Furthermore, the machine learning model will likely aid researchers to classify the large ornithopod dinosaur footprint ichnotaxa, thereby safeguarding against the oversplitting problem.
自 19 世纪以来，除南极洲以外，所有大陆都已确认存在大型鸟脚类恐龙足迹。然而，这些足迹在历史上一直存在种系发生过度划分的问题。为了解决这些问题，并区分已验证的足迹类型，本研究采用基于卷积神经网络的 Xception 迁移学习，自动对鸟脚类恐龙足迹进行分类。机器学习模型使用 274 张数据图像（不包括水平翻转图像），经过 162 个周期（即模型通过所有训练数据的完整周期数）进行训练。训练后的模型准确率为 96.36%，验证准确率为 92.59%。我们使用未包含在训练数据集中的足迹插图来演示机器学习模型的性能。这些结果表明，本研究中开发的机器学习模型可以正确分类大型鸟脚类恐龙足迹插图数据。然而，足迹插图数据（或图像）的质量本质上会影响我们机器学习模型的性能，对于保存完好的足迹，模型的性能更好。此外，由于所开发的机器学习模型是一种典型的监督学习模型，因此无法引入新的标签或类别。尽管本研究使用的是插图而不是照片或 3D 数据，但它是首次在学术层面上应用机器学习技术来验证大型鸟脚类恐龙足迹的种系发生分类。此外，机器学习模型可能有助于研究人员对大型鸟脚类恐龙足迹的分类，从而防止过度划分的问题。
