Krishimitr An IoT Platform for Disease Detection in Agriculture using Convolutional Neural Networks

Abstract

Increasing population of the world is putting more and more stress on agriculture, providing impetus for development of better crops with higher yields. newlineHowever, even with better crop varieties, water and planting management, crop newlinediseases pose a major threat to agricultuists. This threat is especially severe for newlinethe millions of financially stressed small landholder families where crop diseases newlinecan be devastating and timely detection of diseases and proper guidance hard to newlineget. In developing countries like India, farmers often are under significant debts newlineand get by on meager incomes while being the source of food for billions of people. newlineLosing even 10% of their crops to diseases can be the difference between earning a living or going under. This thesis describes work performed to develop a newlinecheap device and framework for an accurate system which farmers can use to stay newlineinformed about disease spread using visible symptoms on leaves. newlineChapter 4 describes an algorithm, which uses machine learning to detect diseases in a wide variety of plants and diseases. High accuracy (gt93%) was obtained newlinewith very noisy images, different backgrounds and different disease coverage. The newlinealgorithm is able to train itself, which means that the accuracy can increase with newlineusage. It can run on a variety of platforms including smartphones and can thus newlineaid non-expert farmers manage diseases effectively. newlineChapter 5 describes a framework utilizing Machine Learning, Cloud Computing and Internet-of-Things, which brings experts to farmers, allowing for timely newlinedetection of diseases. This innovative and comprehensive framework provides newlineagronomists and farmers with a solution for diagnosing plant diseases. By leveraging modern ICT capabilities, this extensible framework is currently trained for newlineover 15 plant types and more than 51 disease types. Our framework employs a hybrid model combining use of both online and offline resources to provide up-to-date newlineinformation to farmers even in case of patchy connectivity.