Identification of broad-leaf trees using deep learning based on field photographs of multiple leaves

We used deep learning to identify broad-leaved tree species based on photographs taken outdoors that included multiple leaves. We took 300 photographs of each of 12 broad-leaved tree species, and used these to produce 43,200 256 × 256 pixels images. We used Caffe as the deep learning framework, and AlexNet and GoogLeNet as the deep learning algorithms. We trained four learning models from a combination of test data ratios, with/without data augmentation, and assessed the classification accuracy based on the ratio of correct classifications to total classifications (“correct ratio”). We divided leaf images into 10 equal sets and conducted 10 iterations without duplication. The results indicated that the highest average correct ratio for the 10 sets was 97.0% in GoogLeNet after 200 epochs. GoogLeNet showed a higher average correct ratio than AlexNet for all of the learning models. A higher proportion of test data resulted in lower average correct ratios without data augmentation. However, the correct ratio tended to improve with data augmentation. Pairs of mutually misclassified species were not necessarily misidentified within categories based on morphological similarity.