AUTOMATIC IDENTIFICATION OF PLANT SPECIES THROUGH A CONVOLUTIONAL NEURAL NETWORK MODEL FOR UAV MOUNTED DIGITAL CAMERAS

摘要

Species Identification is a significant part of biodiversity conservation and protection. Traditional techniques of plant species identification are slow, complicated and require expertise in the field of biosciences. Plant species identification is a challenging task for a novice, interested in obtaining knowledge for various applications such as Bio-diversity monitoring, remote sensing. Hence, with the advent of cost-effective unmanned aerial vehicle technologies, deep learning and computer-vision have given rise to an interest in their use for plant species identification, with minimum knowledge of the expert. In this study, a seven-layer convolutional neural network (CNN) based on deep learning feedforward artificial neural networks is proposed for automatic identification of plant species from the images or videos acquired by UAVs, using feature learning in real time. Small UAVs are suited for this model implementation as they can capture data at a very low altitude. Three types of plant species leaf's i.e. Eucalyptus, Corylus (Hazel), Maple (Acer) where used for training of the network (90% of acquired data) and for testing (10% of acquired data) for identification by the networks. Model performance and efficiency are studied using the accuracy, loss curves and confusion matrix. The model showed an outstanding performance of 93% recognition rate. Performance of model is notable because only basic RGB images are used. It is observed that the increase in training data increases the accuracy of identification with decreased loss rate. The model could be trained to recognize as many species with basic RGB images or videos, without the need for developing a new system. Future work include increase in robustness of the model by training it with a greater number of species. This model can be a powerful tool for automated identification of plant species in very low altitude UAV imageries, videos and could be used for many forest, agricultural research and management processes.