The problem HealthyHarvest solves

Plant diseases pose a significant challenge to agricultural productivity and food security. The inability to detect and manage plant diseases in a timely and accurate manner leads to substantial crop losses, reduced yield quality, and increased economic burden on farmers. Existing methods of disease detection often rely on visual inspection, which can be subjective and prone to human error. Additionally, the lack of accessible and affordable disease management solutions exacerbates the problem, particularly for small-scale farmers and resource-constrained regions.

Our goal is to revolutionize plant disease management practices by leveraging the power of Computer Vision and Machine Learning. Through our innovative solution, we propose to implement a cutting-edge plant disease detection algorithm onto a drone. By deploying the drone above farms, we can efficiently and accurately detect diseases in plants and take immediate measures to combat them. This approach combines the advantages of aerial surveillance, advanced algorithms, and real-time data analysis to provide farmers with timely and actionable insights for effective disease management. Our solution aims to enhance crop health, minimize losses, and optimize agricultural practices using state-of-the-art technology.

Our Computer Vision model of Disease Detection yields and accuracy of 94.97% and a loss of just 0.1525. We have also tried an innovative approach of disease detection using YOLO.

Challenges we ran into

To develop our disease detection model, we faced the challenge of accurately localizing the plant leaves. This necessitated the creation of an additional object detection model capable of identifying and delineating the plant leaves. Building a comprehensive dataset for this purpose proved to be a demanding task, and training the YOLOv5 model for object detection posed its own set of challenges. Achieving an accuracy level above 90% proved to be particularly arduous. Due to the time-intensive nature of training, we were only able to run the model for 10 epochs in each iteration. Despite these obstacles, we persisted in refining our approach to ensure the best possible outcomes.