A novel transfer learning-based approach for plant species prediction using leaf images

There is extensive research going on in the field of automatic identification of plants in order to preserve the plant species which are on the verge of extinction and also to educate people or new generation farmers about the various plants growing in their vicinity. Few plants also possess certain medicinal properties which can be used by the layman to treat commonly occurring ailments. We suggest a novel leaf identification approach using combination of deep learning and conventional machine learning techniques. In this approach, the leaf image features are extracted using a neural network pre-trained on the ImageNet and then fed into the machine learning classifiers for predictions. We prepared three different models and analyzed their performance. Thereafter, we propose an ensemble approach based on stacking classifiers where the predictions of multiple classifiers were used to train a meta-classifier. This approach achieved an accuracy of 99.16% and 98.13% on the unseen samples of Swedish and Flavia datasets respectively.