Design and Development of Smart System for Agricultural Application Using Internet of Things and Machine Learning Techniques

Abstract

newline Farming and Agriculture accounts for a major portion of GDP (Gross Domestic Product) not newlineonly in developing countries but also in many developed nations. As the global population is newlineprojected to reach around 9.6 billion by 2050, improvising and optimizing the current farming newlinetechnologies is the need of the hour. Plant health is an important factor in agricultural production as newlineit mostly affected by plant diseases. Traditionally, plant disease detection has been carried out newlinethrough visual inspection by human experts. This method is based on subjective perception hence it newlinehas risk for error in detecting accurate disease. In recent past, researchers have proposed numerous newlinemachine learning approaches to detect the plant diseases. In this research work, the plant disease newlinedetection system using MobileNetV2 algorithm has been designed and trained with 500 Tomato newlineplant leaf images of five different classes. The edge device has been developed by deploying trained newlinemodel in Sipeed Maixduino development board which is able to detect disease in camera captured newlineinput images. newlineDue to advancement in artificial intelligence and electronic gadgets technology, there is large newlinescope for improvement in neural network algorithms for detecting plant diseases early and newlineaccurately by extracting leaves features efficiently. To detect tomato plant diseases, the novel newlineconvolutional neural network (CNN) model has been designed with hierarchical mixed pooling newlinetechnique using smoothing to sharpening approach. The model has been trained with 10000 images newlinewhich includes 1000 images of healthy leaf and 1000 images each for nine different diseases newlinefrequently occurs in Tomato plant. The different training models has been framed and experimented newlineto identify efficient hierarchy of pooling techniques. The CNN training model exhibit smoothing to newlinesharpening approach with Average-Max-GlobalMax mixed pooling hierarchy and depicts better newlineperformance with a training loss 28.88%, a validation loss 12.61%, a training accu