Leaf species and disease classification using multiscale parallel deep CNN architecture

Plant species are often affected by conquering biotic strains and for sustainable yield more emphasis can be on the novel mitigation measures rather than traditional methods. Plant diseases are witnessed by visible effect on the leaf like the detectable change in color, texture or shape. Categorizing leaf diseases poses challenges like intensity of the disease in the leaf, resolution of the image, shot category and complex background. Literature reports myriads of architecture employing Convolutional Neural Networks for generating models that assist in detecting plant disease. This research work has merged responses from customized filters (Law’s Mask) that well define the texture pattern and learnable filters to ensure adaptive learning. Depending upon the stages of diseases in leaves, the defects occur at varying scales and at varying locations of leaves. Thus, rather than single deep stream of network, a specialized parallel multiscale stream with learnable filters that extract inherent attributes are utilized for improved performance. Experimental evaluation of the proposed methodology with end to end training on Plant Village dataset with 39 classes gives 99.17% for plant species classification and 98.61% for disease classification. For data Repository of Leaf Images with 12 species, 97.16% for plant species classification and 90.02% for leaf disease classification. MepcoTropicLeaf an Indian Ayurvedic Leaf dataset with 50 species is experimented using the proposed algorithm and reported with 90.86% of classification accuracy.