The work presented in this thesis focuses on improving the anomaly detection process for remotely sensed hyperspectral imagery. This process is split up into three main sections; data reduction, atmospheric correction and anomaly detection. The final stage of this anomaly detection process is the actual anomaly detection algorithm. The initial contribution looks at developing a new type of anomaly detection algorithm based on the Percentage Occupancy Hit-or-Miss Transform. Also, a process for trying to improve the existing Mahalanobis Distance technique for hyperspectral data is explained. Both techniques are then tested on two aerial hyperspectral images, and the results are compared with an established technique the Sequential Maximum Angle Convex Cone algorithm. One of the preprocessing steps of the anomaly detection process is the atmospheric correction phase. In this thesis an interface is developed in MATLAB for the atmospheric modelling software MODTRAN, this interface is then used to find the key parameters that have the most effect on the atmospheric models produced. Having determined the key parameters of a MODTRAN atmospheric model, the models are then used to atmospherically correct eight hyperspectral images; four visible to near-infrared and four short wave infrared hyperspectral images. Two scene based approaches for atmospheric correction are also proposed that use known spectra extracted from the scene to produce an atmospheric transform. All three techniques are then evaluated against existing scene-based approaches, namely Internal Average Relative Reflectance and Dark Object Subtraction.The final contribution focuses on the data reduction phase, images of a wind turbine blade with simulated erosion were taken using a near-infrared hyperspectral camera. By analysing the images produced it was possible to determine the optimal bands necessary to detect each type of erosion. The greyscale images produced for the optimal bands were then compared with standard RGB camera imaging to determine if any more detail was shown in the hyperspectral images. Also by imaging the blade at varying light levels, it was possible to determine when this technique breaks down, however by performing some post-processing of the new data using a polynomial surface subtraction to flatten the images it was again possible to extract additional information from the hyperspectral images.
|Date of Award||1 Apr 2017|
- University Of Strathclyde
|Supervisor||Stephen Marshall (Supervisor) & Alison Gray (Supervisor)|