Disease Identification in Leafy Vegetables Using Transfer Learning

Abstract

Abstract: Plants are the major source which gives foods for human to survive. In developing countries like SriLanka agriculture plays a major role in the economic success of people live there and as well as for the whole country's success. In such a situation diseases cause huge losses to farmers. The key concept of maintaining quality and quantity of crops is to detect diseases in earlier stages at the correct time and to take preventive actions against the disease. Usually, farmers recognize diseases through naked eye observation. So, it may not the right caption and it tends to spread wrong pesticides and overdosages of pesticides. Hiring expertise in this area is highly costing and not possible to find that many experts. Here include many techniques used to identify diseases in various types of plants. But those papers do not address the area of Sri Lankan leafy vegetable disease identification. This research work proposed a system with a learning approach for disease identification procedure named transfer learning and fine-tuning, partially tested, and obtain better results. InceptionV3 and VGG16 are the two pre-trained models use to retrain the model. InceptionV3 gain 0.95 training accuracy and 0.79 validation accuracy. VGG16 gain 0.91 training accuracy and 0.86 validation accuracy. At the initial stage the tested system has capable of recognizing brown spot disease at 0.43 and 0.48 testing probabilities in Gotukola, and leaf-spot disease at 0.58 and 0.90 testing probabilities in the Mukunuwanna plant through VGG16 and InceptionV3 respectively.