Determination of suitable image capturing approaches to develop a machine learning application for accurate diagnosis of cinnamon leaf spot disease through teachable machine

Abstract

Accurate and timely identification of pests and diseases is critical for effective pest and disease management; however, information and expert consultations may not be readily available to farmers and may be expensive. Mobile phone technology with cloud computing Artificial intelligence disciplines have advanced significantly over the years, making mobile phones with cameras widely available to people at lower costs and easily usable for identifying plant pests and diseases, surpassing the constraints of conventional techniques. The effect of using mobile phone cameras and phone camera-attached lenses to train an Artificial Intelligence (AI) engine to identify cinnamon leaf spot disease was experimented under several conditions. Three smartphone cameras (64 MP, 48 MP, and 8 MP), two camera attached lenses (10x and 30x magnifications), with or without flashlights, and two sides of the leaf (upper and lower) were taken as different conditions to make 24 treatment combinations. Fifty images of diseased leaves and 50 images of healthy leaves were obtained under each combination and image processing engines were trained for each combination by the open-source application called “Teachable Machine” by uploading images of diseased leaves and healthy leaves for each class. Then engines developed were validated with 10 healthy and 10 images with diseased leaves captured from a 9.5 MP camera under the same treatment combinations. Results revealed that the quality of the camera, AI lens, and flashlight conditions used to take the images did not affect the accuracy of identification by the engine. The trained engines could be deployed to develop a mobile-based disease diagnosis app for field use.