Deep learning for improved plant classification

Abstract

Plants are crucial in the ecosystem: they enable life by providing oxygen, they have medicinal properties, among many other uses, and hence their classification is vital. Automating plant identification has remained a challenging task. Machine learning does not automate feature engineering, making its application to the problem an arduous task. Deep learning automates featuring engineering, but requires large datasets, presenting another challenge. This research proposes an improved and fast plant classification model for plants using deep learning on small datasets. The objectives of this research are to explore techniques that are state-of-the-art in the classification of plants using leaves, to improve the classification accuracy scores, and to model a framework for the classification. A convolutional neural network was designed and applied on a small leaf dataset with fine-tuning. An accuracy score of 94.99% was achieved. High overfitting was noted since the dataset was small. Data augmentation was applied to the dataset with the images augmented before being input into the model, differing from the many cases where augmentation is applied on-the-fly. This increased the success rate to 99.99% with reduced overfitting. This showed that augmentation applied in this way improves the performance of the model. Transfer learning with the same augmentation method was applied, resulting in a 100% test accuracy and stable results with low overfitting. The proposed methodology provided good results on the Flavia leaf data set used in the experiments. Finally, a deep learning framework for improved plant classification using leaves is outlined.