Tree-level identification and classification leveraging image data analysis

Abstract

The environmental monitoring, agriculture and forestry management depend on the categorization of trees like fruit, medicinal or forest tree. The aim of this project is to develop a trustworthy deep learning system to recognise these classifications of trees based upon images of their leaves and bark. Thus three DNN architectures were used: ResNet50, ResNet101 and InceptionV3. The dataset used for training and testing included a diverse set of images taken from different sources, and showing different species of trees in different environments. These images were preprocessed, resized, normalized and augmented so as to ensure that the models can learn and generalize effectively. They are transfer learned using pre-trained ImageNet weights for the three models. We evaluated the models based on performance metrics (accuracy and loss) on both training and test sets. The ResNet101 is only one of the 3 models trained with a maximum test accuracy of 85%. ResNet50 and InceptionV3 produced similar results with slightly lower accuracy. This result also gives an evidence to indicate that the deep’s ability of deeply structure such as ResNet101 on merging the complex characteristics and patterns over leaf and bark images for tree classification. The results in this paper can provide basis for practical applications of automatic tree identification, e.g., forest management and ecologic studies. In the future, the methodology and research can be expanded by involving a larger data set and multiple models to enhance both accuracy and scalability.