Hybrid CNN Models for Plant Species Recognition and Disease Detection

Plants and trees play a crucial role to maintain life on earth, as it is the initial component of the food chain and hence they are inevitable for life sustainability. Human activities have resulted in increased plant diseases and endangered the habitats essential for the survival of a large number of plant species. Conservation of plant species and its disease control has become an imperative and emergency need to recover, restore, and maintain the plant diversity of natural habitats. Rapid developments in computer vision and mobile communication devices can greatly benefit the automation of plant species recognition, plant disease identification, and disease control. The proposed model is a combined system for disease detection and species identification by using images of healthy and unhealthy leaves. The model is trained and validated in an integrated dataset created by combining the standard datasets of species identification and disease detection of medicinal leaves. Proposed work studies the performance of two different hybrid deep learning architectures. First one makes use of Convolutional Siamese Network (CSN) a twin CNN model which requires only a small number of training samples. Second uses CNN followed by ensemble classifier configuration. Two types of base CNN models used are simple basic model (SM) and pre-trained model VGG16. The experimentation has shown that using a multi-learners model, CNN accuracy can be improved with fewer training images, without data augmentation, with less feature dimensionality. On an average 13.4% and 37.39% increase in accuracy was observed while using multi-learner model with CNN VGG16 model and CNN simple model (SM). CSN performance is better compare to that of CNN and CSN performs well with simple model compared with VGG16 and multi-learners.