A Deep Convolutional Neural Network Approach for Plant Leaf Segmentation and Disease Classification in Smart Agriculture

This paper concerns an approach for developing a set of tools based on computer vision and deep learning to detect and classify plant diseases in smart agriculture. The main reason that motivated this work is that early detection of plant diseases can help farmers effectively monitor the health of their culture, as well as make the best decision to avoid the spread of the pathogens. In this work, a novel way for training and building a fast and extensible solution to detect plant diseases with images and a convolutional neural network is described. The development of this methodology is achieved in two main steps. The first one introduces Mask R-CNN to give bounding boxes and masks over the area of plant leaves in images. The model is trained on the PlantDoc dataset made of labeled images of leaves with their corresponding bounding boxes. And the second one presents a convolutional neural network that returns the class of the plant. This CNN is trained on the PlantVillage dataset to recognize 38 classes across 14 plant species. Experimental results of the proposed approach show an average accuracy of 76% for leaf segmentation and 83% for disease classification.