Design and Evaluation of Classification Models Using Machine Learning Approach for Plant Diseases

Abstract

Plants are regarded as an essential source of life on the Earth. Visual investigation of any plant disease by naked eyes is still the primary technique of disease diagnosis in villages of developing countries. Generally, this self-investigation does not capture initial symptoms of the infection at early stage of diseases which gives chance to the pathogens to spread out in the whole crop fields. Many times, this visual inspection fails to recognize the diseases accurately. Late diagnosis causes a reduction in crop yields. Hence, automated and accurate disease identification in plants is very important to ensure better quantity and quality of crops. Automatic-detection of crop illnesses is a key research domain because it might help in monitoring the huge fields of crops and, as a result, recognize disease signs as-soon-as they occur on plant leaves. newlineVarious relevant studies related to plant disease diagnosis and classification is selected for literature review of this research. We have identified following research gaps during the literature review. In this we have done the Review of the literatures to find the limitations of the most effective methods utilized in plant leaf disease detection and severity estimation. newlineNext, we have explored whether a simple fully connected Convolution Neural Network architecture can be designed and utilized to recognize and classify the rice plant illnesses effectively by using images of plant leaf, and whether a background removal technique can be utilized to increase the performance even more. For exploration, we have developed a simple fully connected CNN for rice illness (pathogen) recognition and classification using leaf images. Number of filters and size of filters in convolution operation are designed in such a good manner that it outperformed the existing methods with an accuracy of 99.1%. We have used large dataset to overcome the problem of overfitting. To boost performance even more, we have applied background removal technique on input images before feeding it into