A Mobile-Based Novice Agriculturalist Plant Care Support System: Classifying Plant Diseases using Deep Learning

Abstract

For novice gardeners and small-scale farmers, identifying the exact diseases that plague their plants and determining the best treatments for the species affected can be a difficult task without the correct expert knowledge, yet is one that is crucial to sustain successful plant growth. Our project aims to assist gardeners who do not have the required expert knowledge about the characteristics of species-and-disease combinations to avoid the major pitfalls of misidentifying or mistreating diseases by developing a plant care support system containing a high-accuracy, multi-label classification model to classify the [species-disease] combination of a plant non-invasively from an image of the plant’s leaf. The project also aims to provide further agricultural guidance to gardeners by outlining the appropriate plant care details for identifying, treating, and preventing the classified [species-disease] combinations, helping gardeners grow their field knowledge. Our work consists of a brief comparative analysis between lite, mobile-optimized CNN models to attempt to maximize the accuracy and performance of our classification solution, building on pre-trained base networks using transfer learning. We investigate the effects of varying the retrained portions of the base networks, the effects of using different CNN architectures, and the effects of varying the network hyperparameters on the models’ performances. With these objectives, 32 model variations were developed and evaluated using various standard metrics including accuracy, F1-score, and confusion matrices. The best performing model was found to be an EfficientNetB0 model using a fully retrained base network with optimized hyperparameters, and was then integrated into a system composed of a frontend mobile application and backend centralized cloud database. Beyond the classification and plant care support functionalities, the project also aims to generate new spatiotemporal analytics about the common global species-disease trends by region and season using the collective users’ classification results, making these analytics available to all system users and contributing to the efforts of better understanding global agricultural trends.