Banana Seedlings Health Monitoring For Micro Air Vehicles Using Deep Convolutional Neural Network

• Main Author: Tan, Shu Chuan
• Format: Monograph
• Language: English
• Published: Universiti Sains Malaysia 2021
• Subjects: T Technology
• Online Access: http://eprints.usm.my/54694/1/Banana%20Seedlings%20Health%20Monitoring%20For%20Micro%20Air%20Vehicles%20Using%20Deep%20Convolutional%20Neural%20Network.pdf; http://eprints.usm.my/54694/

Banana is one of the most produced but also highly demanded fruits in Malaysia. Many farmers invested in tissue-cultured techniques in greenhouses to increase production, but the tissue-cultured banana seedlings are not invincible to numerous diseases and pest attacks. To monitor the health conditions of the tissue-cultured banana seedlings, they need to hire many laborers or install cameras or sensors throughout the greenhouse. To this end, we proposed an affordable, hand-palm-sized, and automatic Micro Air Vehicle (MAV) to help farmers. MAVs have the mobility to access every corner of confined spaces and acquire an excellent bird’s-eye view, which provides great convenience in monitoring tasks, and is capable to fly according to the desired flight path and capture each plant precisely. This research compares the performances of five YOLO and Single Shot MultiBox Detector (SSD) deep learning model architectures in predicting the health status of banana seedlings. The Tiny-YOLOv4 model architecture, which has the best compromise between detection accuracy and detection speed, was then trained with different network resolutions and weightage of negative samples. Tiny-YOLOv4 with the 416×416 network resolution and 18% of negative samples has the highest mAP of 99.08% and was chosen to categorise the plants into normal and unhealthy classes based on the images captured by an onboard camera. Several flight tests were performed successfully in an indoor hall, and the plants were classified accurately. The locations of unhealthy plants are sent to notify farmers of further actions. The proposed solution in this project is expected to highly reduce labor-intensive activities and possible human error.