The development of methodology and techniques for crop disease identification

Abstract

In India, an estimated 15-25% of potential crop production is lost due pest and diseases (Roy and Bezbaruah, 2002). The country needs not only to raise production but also ensure food security for its growing consumption needs while curbing excessive pesticide usage. Detection of pests and diseases at an early stage plays a significant role in addressing the above-mentioned concerns and image classification offers a cost-effective and scalable solution to the disease detection problem (A. Ramcharan et al. 2017). Here, the principles of transfer learning are implemented with pretrained model – Resnet34 (K. He et al. 2015), and test its effectiveness in image classification using a dataset of tea leaves. The novelty of this work is that the images used are not curated, individual leaves with controlled backgrounds but of plants in-situ. The effect of the level of zoom and background is examined and class activation maps are used to validate that the basis of classification is indeed the disease and not an artificial bias from factors such as background, lighting etc.