A Transfer Learning Classification Prototype Model for Sri Lankan Trees

Abstract

Automatic classification of the trees from images or videos is difficult, due to its large variety of species. Identifying and classifying tree types requires agricultural experts and botanists. The technology is moving fast so the people are running towards growth so they become more urbanized and forget about nature. We have lost valuable tree species due to the forest firing without knowing its name. It’s difficult to identify trees by the naked eye because of their variation in orientation, the shape of the leaves, and the viewpoint. Next-generation people mostly will forget the importance of the trees. So there is a need for a system to recognize trees. In this experiment, we acquired 2,500 images of 25 different types of trees inside South Eastern University of Sri Lanka premises. We focused on analyzing the performance of the pre-trained deep learning models. Therefore, we specifically use transfer learning by leveraging pre-trained deep learning models. The evaluated architectures are VGG16, VGG19, Xception, InceptionV3, and MobileNet. The performance of each classification model is assessed in terms of validation accuracy. In our experiment, MobileNet achieves a validation accuracy score of 87.20% and outperforms other pre-trained models. Trees must be classified into identifiable classes for us to have a clear, organized way of identifying the diverse array of trees that inhabit the planet. We did not find any similar research work on identifying/ classifying trees in Sri Lanka using image processing-based machine learning. This MobileNet architecture could be used to deploy the final web-based application solution and mobile-based application. Also, it took a smaller number of epochs which is 4 to reach the higher validation accuracy. Because of the lightweight nature of the MobileNet architecture model, it has performed better than others.