TinyML for the Detection of Plant Diseases in Resource-Constrained Areas within West Africa

Most of West Africa's rural farmers do not have access to smartphones, a stable internet connection or stable electricity - additionally, they often have low English proficiency and technological competency. Subsistence farmers in the Akuapim-South Municipal Assembly of Ghana face the same reality but have the added problem of limited access to agricultural extension agents who provide plant disease identification services. While many technological solutions to plant disease identification exist, they are cloud-based and, thus, require resources not accessible to a rural farmer in this population. This study investigates the use of Tiny Machine Learning (TinyML), a subset of machine learning that is focused on deployment to microcontrollers and hence facilitates low latency, low power, low cost, and high privacy computing. While this technology has been proven comparable to cloud-based Machine Learning, there is a lack of research regarding its applicability to both the area of plant disease identification and the West African context. To fill this gap, this study develops a TinyML-based system and compares it to an existing cloud-based plant disease identification mobile app. The results of this study suggest TinyML has promising performance although it falls short of the performance of the cloud-based solution in this study. A significant advantage of the TinyML system is on-device inference without the need for internet connectivity. Going forward, a combination of cloud-based model training and on-device inference is suggested as the most suitable ML setup to address this problem for this population. This study also identifies some areas for improvement and further research, such as the creation of a context-specific dataset, expanded experiment scope, and the use of alternative TinyML development platforms.