Early Detection Of Oak Wilt Using Machine Learning And Unmanned Aerial Vehicles (Uavs)
Location
Hager-Lubbers Exhibition Hall
Description
Oak wilt, a highly contagious and lethal disease threatening oak trees across North America, has traditionally been detected through manual inspections, which are often time-intensive and insufficient for timely intervention. This study introduces an automated detection system leveraging aerial imagery captured by unmanned aerial vehicles (UAVs) and a convolutional neural network (CNN) to classify environments as "Oak Wilt" or "Not Oak Wilt." Initially trained on a dataset of 581 images, the model was further refined with an expanded dataset of 1,051 images and tested on 9,981 images collected from various state parks in Michigan. The customized CNN architecture, which includes convolutional, max-pooling, and fully connected layers, achieved an overall accuracy of 86.72%. The system incorporates data augmentation, geotagging, and Reinforcement Learning from Human Feedback (RLHF), significantly enhancing model adaptability and accuracy. Deployment through a web platform built with VueJS and Flask improves accessibility and scalability, facilitating proactive forest conservation by enabling early oak wilt detection from aerial imagery.
Early Detection Of Oak Wilt Using Machine Learning And Unmanned Aerial Vehicles (Uavs)
Hager-Lubbers Exhibition Hall
Oak wilt, a highly contagious and lethal disease threatening oak trees across North America, has traditionally been detected through manual inspections, which are often time-intensive and insufficient for timely intervention. This study introduces an automated detection system leveraging aerial imagery captured by unmanned aerial vehicles (UAVs) and a convolutional neural network (CNN) to classify environments as "Oak Wilt" or "Not Oak Wilt." Initially trained on a dataset of 581 images, the model was further refined with an expanded dataset of 1,051 images and tested on 9,981 images collected from various state parks in Michigan. The customized CNN architecture, which includes convolutional, max-pooling, and fully connected layers, achieved an overall accuracy of 86.72%. The system incorporates data augmentation, geotagging, and Reinforcement Learning from Human Feedback (RLHF), significantly enhancing model adaptability and accuracy. Deployment through a web platform built with VueJS and Flask improves accessibility and scalability, facilitating proactive forest conservation by enabling early oak wilt detection from aerial imagery.