基于树木年轮径切特征的卷积神经网络树种识别

卷积神经网络可以通过树木年轮样本构造特征图像实现物种识别的自动化。本研究通过建立树木年轮样本构造特征图像集,选用LeNet、AlexNet、GoogLeNet和VGGNet 4个卷积神经网络模型,实现基于树木年轮横切面的计算机自动化树种精准识别,进而确定各模型的树种识别准确率,明晰不同树种在自动识别中的混淆情况,探测不同模型识别结果的差异。结果表明: 本研究训练的用于树种识别的卷积神经网络模型具有较好的可信度;4个模型中GoogLeNet模型树种识别准确率最高,为96.7%,LeNet模型识别准确率最低(66.4%);不同模型对于所选树种的识别结果具有一致性,表现为对蒙古栎识别准确率最高(AlexNet模型识别率达到100%),对臭冷杉的识别准确率最低。本研究中也存在类似结构树种的识别混淆情况。模型在科和属水平的识别准确率高于种水平;阔叶树种因其显著的结构差异容易区分,阔叶树树种的识别准确率高于针叶树。总体上,通过卷积神经网络,探测了树木年轮特征的深层信息,达到树种的精准识别,提供了一种快速便捷的自动树种初筛鉴定方法。

Convolutional neural network tree species identification based on tree-ring radial section image features

Convolutional neural networks can automatically identify tree species based on the images of structural features of tree-rings samples. In this study, we used a tree-ring image dataset for different species to achieve tree-ring based species automatic identification with high accuracy by four convolutional neural network models (LeNet, AlexNet, GoogLeNet, and VGGNet), aiming to determine the identification accuracy of the models, clarify the species misidentification during the automatic processes, and explore the identification differences among the models. The results showed that tree species identification derived from the trained convolutional neural network models was reliable, with the GoogLeNet and LeNet showed the highest (96.7%) and lowest (66.4%) identification accuracy. The tree species identifications using different models were highly consistent. Quercus mongolica and Abies nephrolepis showed the highest (100% for AlexNet) and lowest identification accuracy, respectively. Misidentification could occur among tree species with similar tree-ring structure. The identification accuracy of the models was higher at family and genus levels than that at the species level. The identification accuracy of broadleaf tree species was higher than that of coniferous trees due to distinct radial structure among broadleaf species. Overall, our method achieved a high accuracy for tree species identification, and provided a fast, convenient, and automatic tree species identification by detecting specific tree-ring properties with convolutional neural network.