Identification of varieties in Camellia oleifera leaf based on deep learning technology
Zhipeng Dong, Fan Yang, Jiayi Du, Kailiang Wang, Leyan Lv, Wei Long
Abstract
Camellia oleifera, a woody oil tree, is widely recognized for its valuable oil. Different cultivars of C.oleifera exhibit distinct growth characteristics, oil content, and oil composition. Therefore, the classification of C.oleifera cultivars can aid in the better utilization of C.oleifera resources and improve yield and quality. However, the identification of cultivars remains challenging due to genetic diversity, similarities in leaf morphology, and the influence of geographical environment, among other factors. Comprehensive cultivar identification methods for studying C.oleifera must be established to overcome these obstacles. We selected 118 varieties that grew under natural light conditions and collected whole pest-free mature leaves. After filtering out invalid images, we constructed a leaf cultivar dataset consisting of 30890 images of C.oleifera. The results showed that RegNetY-4.0GF-Convolutional Block Attention Module provides significant advantages over other methods in cultivar recognition, including VGG16, ResNet50, EffificientNet-B4, and EffificientNet-B4-CBAM. It achieved an overall accuracy of 93.7 % and an F1-score of 0.945, much higher than the accuracy of other compared methods. CBAM can significantly improve the accuracy of varieties recognized. The overall results showed that deep learning could effectively distinguish C.oleiera leaves of different varieties. This method provided an effective way to identify C.oleifera varieties quickly and nondestructively.