Plant classification using convolutional neural networks

Abstract

Application of the benefits of modern computing technology to improve the efficiency of agricultural fields is inevitable with growing concerns about increasing world population and limited food resources. Computing technology is crucial not only to industries related to food production but also to environmentalists and other related authorities. It is expected to increase the productivity, contribute to a better understanding of the relationship between environmental factors and healthy crops, reduce the labor costs for farmers and increase the operation speed and accuracy. Implementing machine learning methods such as deep neural networks on agricultural data has gained immense attention in recent years. One of the most important problems is automatic classification of plant species based on their types. Automatic plant type identification process could offer a great help for application of pesticides, fertilization and harvesting of different species on-time in order to improve the production processes of food and drug industries. In this paper, we propose a Convolutional Neural Network (CNN) architecture to classify the type of plants from the image sequences collected from smart agro-stations. First challenges introduced by illumination changes and deblurring are eliminated with some preprocessing steps. Following the preprocessing step, Convolutional Neural Network architecture is employed to extract the features of images. The construction of the CNN architecture and the depth of CNN are crucial points that should be emphasized since they affect the recognition capability of the architecture of neural networks. In order to evaluate the performance of the approach proposed in this paper, the results obtained through CNN model are compared with those obtained by employing SVM classifier with different kernels, as well as feature descriptors such as LBP and GIST. The performance of the approach is tested on dataset collected through a government supported project, TARBIL, for which over 1200 agro-stations are placed throughout Turkey. The experimental results on TARBIL dataset confirm that the proposed method is quite effective.