Hyperspectral vision methods for automatic recognition of emergency plant pests

Abstract

The objective of this research work is to develop the computational techniques in computer vision and machine learning to detect and identify Emergency Plant Pests (EPPs) in hyperspectral images and the ability to apply in a real-time basis for the plant biosecurity surveillance application. There are three problems that will be addressed regarding hyperspectral imagery classification for EPPs; feature extraction, object detection and classification problems. Therefore, the development of computational methods for the three problems have been proposed. The first one, extraction of descriptive features, is to relate EPPs to their spectral imaging representations as necessary for representing information of the pest for reliable classification. This involves the development of texture based image descriptors that are both, invariant to changes in the viewing conditions and robust to in-field conditions. The second one is the object detection related problem in hyperspectral images. For any development of automatic surveillance systems, rapid object detection is one of essentials for success. The proposed method rapidly finds potential areas of locations of suspected objects and operates effectively on large-scale clustering and high dimensionality problems like the hyperspectral data. The last major problem of the research is to develop a classification method appropriate for the application. The method requires the capability to operate under a real-time setting with desirable discrimination performance for satisfying with a high margin of confidence. The development of these methods is further explained based on mathematic formulations. The experimental results show the utility and performance of the proposed methods. Finally, an attempt of further investigation of possible applications for these computational techniques has been carried out. -- provided by Candidate.