Implementing deep learning architectures to extract vegetation composition from drone-based imagery to monitor revegetation on a rehabilitated mine site

Remote sensing is a powerful technology that enables the collection of information about the Earth's surface from a distance, generally using sensors installed on satellites, manned aircraft, or drones. Creating tree crown maps through manual delineation and visual interpretation of imagery has a long history in forestry and ecology. When integrated with artificial intelligence (AI), object detection in remote sensing has become a cutting-edge strategy for automating the extraction of meaningful information from remotely acquired data. Machine learning has become one of the most effective classification techniques in the last decade, with numerous methods continually being developed. Deep learning methods with convolution neural networks (CNN) have received the most attention in current image classification literature revolutionising image processing for RGB imagery. There has been limited research that uses high-resolution HSI acquired from a drone for image classification using CNNs, and it is difficult to obtain a robust solution due to the complexities in acquisition and processing. This project will investigate these complications in a systematic manner, with the goal of developing a processing pipeline that can leverage latest state of the art RGB deep learning architectures to be applied to hyperspectral imagery to achieve much greater accuracy in forestry/woodland structural and floristics assessments.