Precise weed and maize classification through convolutional neuronal networks

• Published in: 2017 IEEE Second Ecuador Technical Chapters Meeting (ETCM) pp. 1 - 6
• Main Authors: Andrea, Cordova-Cruzatty, Mauricio Daniel, Barreno Barreno, Barrionuevo Jose Misael, Jacome
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-10-2017
• Subjects: Agriculture; Approximation algorithms; Artificial neural networks; Classification algorithms; Image sampling; Image segmentation; Machine learning; Machine learning algorithms; Real-time systems; Supervised learning; Training
• DOI: 10.1109/ETCM.2017.8247469

Deep Learning is playing an important role in big data processing for more accurate modeling of common productive processes. It is being widely used in artificial vision applications and specifically in pattern recognition. The versatility of deep learning has positioned it as a fit tool used in many fields of application, among which is precision agriculture. This paper presents the development of an algorithm capable of image segmentation and classification. Segmentation is intended to separate the target plant from the original image, while classification is meant to identify what images belong to the two defined classes. It applies a convolutional neural network (CNN) to discriminate maize plants from weeds in real time, at early crop development stages. It was applied to maize crop because it is a common staple crop in the Ecuadorian Highlands. The convolutional neural network has been trained with a dataset generated in the segmentation stage. The performance of the network was analyzed with LeNET, AlexNet, cNET and sNET network architectures. The network architecture that presented the best training results was cNET based on its performance in terms of accuracy and processing time. The minimum working filter number for this network architecture was 16. The best performing algorithms and processors have a significant potential for autonomous weed and crop classification systems in a real-time application.