Identification of hybrids between the Japanese giant salamander and Chinese giant salamander using deep learning and smartphone images

Biological invasions are recognized as one of the factors causing biodiversity loss. Incomplete reproductive isolation with a closely related species can result in hybridization when a non-native species is introduced into a new habitat. Management of hybrids is essential for biodiversity conservation; however, the distinction between the two species becomes a challenge in cases of hybrids with similar characteristics to native species. Although image recognition technology can be a powerful tool for identifying hybrids, studies have yet to utilize deep learning approaches. Hence, this study aimed to identify hybrids between native Japanese giant salamanders (Andrias japonicus) and non-native Chinese giant salamanders (Andrias davidianus) using EfficientNet and smartphone images. We used smartphone images of 11 native individuals (with 5 training and 6 test images) and 20 hybrid individuals (with 5 training and 15 test images). In our experimental environment, an AI model constructed with efficientNet-V2 showed 100% accuracy in identifying hybrids. In addition, highlighting the regions that influenced the AI model's predictions using Grad-CAM revealed that salamander head spots are responsible for correctly classifying native and hybrid species. The results of this study revealed that our approach is one of the methods that enable the identification of hybrids, which was previously considered difficult without identification by the experts. Furthermore, since this study achieved high-performance identification using smartphone images, it is expected to be applied to a wide range of low-cost identification using citizen science.

生物入侵被认为是引发生物多样性丧失的关键诱因之一。当外来物种被引入全新栖息地后，其与近缘物种间存在的不完全生殖隔离，可能引发杂交事件。杂交个体的管理是生物多样性保护的核心环节之一，但当杂交个体与本土物种外观特征高度相似时，二者的物种区分便成为棘手难题。尽管图像识别技术可作为杂交个体鉴定的有力工具，但目前相关研究尚未应用深度学习方法。因此，本研究旨在借助EfficientNet模型与智能手机拍摄的图像，实现本土日本大鲵（Andrias japonicus）与外来中国大鲵（Andrias davidianus）的杂交个体鉴定。本研究使用了11份本土个体的智能手机图像，其中5张作为训练样本、6张作为测试样本；另包含20份杂交个体的智能手机图像，其中5张作为训练样本、15张作为测试样本。在本实验环境下，基于EfficientNet-V2构建的人工智能模型在杂交个体鉴定任务中实现了100%的识别准确率。此外，借助梯度类激活图（Grad-CAM）可视化对人工智能模型预测结果产生影响的图像区域后发现，大鲵头部的斑点特征是实现本土物种与杂交个体精准分类的核心依据。本研究结果表明，所提出的方法可实现此前仅能依靠专家完成的杂交个体识别，为该领域提供了一种可行解决方案。此外，本研究依托智能手机图像实现了高性能的物种鉴定，有望通过公民科学模式推广至低成本、大范围的物种识别场景中。