Plant disease detection using deep deep learning

Abstract

Agriculture plays an important role in the economic growth of a country. It is really essential to increase our food capacity to cope with the increasing hunger of the world, which is increasing day by day due to the increase in the human population. Agriculture is not only a source of food but also provides employment opportunities to a large portion of the population. However, plants are often attacked by various viral, fungal and bacterial diseases which cause major loss to crops as well as the economy. The timely and accurate detection of plant disease helps to take precautionary measures to control the spread of disease as well as its cure. Through this better results is achieved and it would reduce the expense to be incurred on pesticides and fumigation. It will lead to precise ag1iculture in which only the infected/diseased portion among the entire field is sprayed rather than the whole crop. A lot of work has been done in this research area. Researchers have been experimenting with different methodologies. The emergence of deep learning and also rapid growth in fast computing has lead to better accuracy in the detection of plant disease. In Deep Learning, most of the work has been done on the Plant-Village dataset. Few researchers have collected their own data-set but they are not publicly available. They have achieved really good accuracy on this data-set but when such datasets trained models are tested on in the real world scenario, they didn 't perfonn up to the expectations. The major reason is the images in most of the datasets used for model training are not that complex or somehow not according to real-world field images. To get better results on real-world field images, the model is generalized on five different datasets. The datasets used are different in terms of the number of classes, the variety of diseases and mainly in the conditions in which their images are taken. These datasets are Citrus dataset, Plant diseases dataset, Plant-Pathology dataset, Database of leaf images and PlantDoc dataset. Plant disease dataset and Database of leaf images dataset comprised of 12 classes each with 9026 and 4503 no of images respectively. The citrus dataset comprised of 5 classes with 609 no of images. The plant-Pathology dataset contains 4 different classes with 3651 no of images. PlantDoc dataset has 2598 total images with 27 classes of 13 plant species. These datasets also have a class imbalance problem. For example, in the Citrus dataset one class have only 13 no of captured images. Transfer learning is used as a technique for model training and testing on the abovementioned datasets. Trained Efficient model B7 with noisy students as weights is used as a model. Focal loss is used as a loss parameter. To avoid over-fitting, augmentation and label smoothing is used. F-1 score of 96%,98%,99%,92% and 68% is achieved on the desired datasets respectively. This shows that model is performed well on different types of image datasets with different image complexities.