Federated Machine Learning and TinyML Inference for Crop Disease and Pest Classification on Smartphones

Abstract

As the agricultural industry undergoes a technological revolution, the integration of machine learning (ML) and mobile technologies emerges as a promising solution to address crop disease management efficiently. In this thesis, we present a novel approach combining federated learning (FL) and TinyML inference for crop disease classification on smartphones. Our research encompasses the development of a web application for dataset collection, complemented by a mobile application tailored for farmers. Through rigorous training, we produced multiple ML models, each specialized in detecting diseases across different plant types. These models were subsequently hosted for offline use, empowering farmers with real-time disease identification capabilities directly on their smartphones. Leveraging FL techniques, our solution ensures adaptability and scalability, crucial factors in the agricultural domain. Furthermore, employing TinyML inference enables efficient model execution on resource-constrained devices without compromising accuracy. Evaluation results demonstrate an impressive average accuracy of 98% across all deployed models. This framework represents a significant step forward in democratizing access to advanced agricultural technologies, enhancing crop disease management, and contributing to global food security.