Using Machine Learning Models for Image-Based Crop Disease Detection

Abstract

The early and accurate diagnosis of plant diseases is a crucial factor in ensuring agricultural productivity and mitigating both economic and environmental consequences. Traditional disease detection methods often rely on manual inspection, which can be time consuming, prone to human error, and impractical for large-scale agricultural applications. To address these challenges, this thesis focuses on the development and optimization of an automated plant disease detection and classification system leveraging advanced deep learning techniques. Specifically, a method based on Mask R-CNN with Instance Segmentation is proposed, allowing for the precise localization and identification of infected leaf regions. For the system's implementation, Detectron2, a powerful object detection framework, was employed in conjunction with Transfer Learning to improve classification accuracy while reducing the need for extensive labelled datasets. The developed approach is capable of detecting and classifying eight distinct plant diseases, offering an efficient and scalable solution for automated disease monitoring. By leveraging pre-trained models, the system enhances generalization across different plant species and environmental conditions, making it a practical tool for real-world applications. To enable deployment on low-power edge devices, Pruning and Quantization techniques were applied, significantly reducing the model size and computational requirements while maintaining high detection accuracy. Experimental evaluations confirm the effectiveness of the proposed approach, demonstrating that optimized deep learning models can deliver reliable performance even under constrained computational resources. This optimization ensures that the system can be integrated into smart farming applications, handheld diagnostic devices, or embedded systems in agricultural environments.