Portable Apple Leaf Disease Detection System Using Convolutional Neural Networks
A major issue in agriculture is the need to cope with plant diseases. Adequate expertise and domain knowledge are required to handle diseases effectively in the field. The concurrence of the increasingly portable device ownership, such as smartphones and drones, and the major advances in Machine Learning opens a potential measure for portable devices in disease diagnosis at a low cost. This research studies the challenge and solution of building an apple leaf disease detection system using Deep Learning, which can operate in the field without requiring a plant expert presence. The core of the system adopts recent convolutional neural networks, build up two-stage machine learning application pipeline to diagnose apple leaf disease directly from images captured on the field without causing any damage to plant. In the research, transfer learning and fine turning techniques are considered and applied effectively, helping to reduce training time while pre-trained knowledge is reused. Other factors that are dealt with in this system are optimisation, made it feasible for mobile and edge device running at limited resources. Both original and optimised versions of disease detection model are yielding impressive accuracy score which is more than 0.99. The leaf segmentation model is able to work within the system pipeline as providing visually validated results. The entire system should be built and wrapped in a portable, cross-platform application, as well as deployable and serve as cloud services. This research was encouraged and oriented by the design science research process framework .