利用深度卷積類神經網路辨識形態相似的樟屬(樟科)物種

Abstract

樹苗摻假是林業生產上一個現存的問題。土肉桂為常綠植物，由於具有豐富的肉桂醛，因此有較高的經濟價值。和土肉桂具有相似葉部形態的另外兩種樹種：陰香和台灣肉桂，則不富含肉桂醛這種化學成分。在台灣，已有相當多報導指出有關陰香摻假土肉桂的事件。然而，由於它們外觀形態的高度相似性，因此分辨這三種樹種相當不容易。而樹種之間的價值差異，也常造成林業農民的經濟損失。本研究提出利用葉子影像和深度學習來辨識三種樟屬樹種。利用平板掃描器獲取兩個不同地區的葉子影像，然後使用一個地區採集的葉子影像作為訓練樣本，開發了基於VGG16，Inception-V3和NASNet模型的深度卷積神經網絡分類器。結果顯示，這些深度卷積神經網絡分類器應用在另一個地區的葉子影像上，測試準確率達到至少0.87。此外，本研究開發的卷積神經網絡分類器也優於支持向量機分類器。

Tree and seedling adulteration has becoming an issue in plant cultivation. Cinnamomum osmophloeum (Lauraceae) is an evergreen plant that yields cinnamaldehyde compound and has high economic value. Although morphologically resembling C. osmophloeum, two other species, C. burmannii and C. insulari-montanum, do not produce cinnamaldehyde. Adulteration of C. burmannii using C. osmophloeum has been reported in Taiwan. Yet, even for experts, it is challenging to discriminate the three species from their appearance due to their high degree of similarity. This brings economic loss to forest farmers owing to the value discrepancy between the species. This study proposed to identify the three Cinnamomum species using leaf images and deep learning approaches. In the study, leaf images of the three species were acquired from two camps using flatbed scanners. Deep convolutional neural network (CNN) classifiers based on VGG16, Inception-V3, and NASNet models were then developed using the leaf images collected from one garden as the training samples. The result showed that the developed deep CNN classifiers reached a test accuracy of at least 0.87 on the images collected from the other garden. The developed CNN classifiers also outperformed support vector machine classifiers.