Degree Name

Doctor of Philosophy


School of Electrical, Computer and Telecommunications Engineering


Ground penetrating radar has been widely used in many applications, such as archaeological explorations, glacier and ice sheet investigation, sedimentological research, paleolimnology studies, detection and monitoring of below-ground biological structures, mineral exploration and resource evaluation, building condition assessment, road pavement and bridge deck analysis, and landmine detection. However, the processing and interpretation of the acquired signals remain challenging tasks.

This dissertation focuses on an automatic classification system for GPR traces that minimises human intervention. In a GPR survey, particular resonance frequencies arise in wave propagation; therefore, reflected waves from different buried objects or paths present different electromagnetic characteristics. Inspired by these observations, three different approaches are proposed for the classification of railway ballast fouling conditions and evaluated on real-world railway GPR data.

The first approach classifies the buried objects or underground materials by analysing the frequency spectra of the received GPR signals. The proposed system extracts features from magnitude spectrum using the morphological dilation and categorizes the features through support vector machines. The experimental results indicate that the proposed salient spectrum amplitudes are an efficient representation of ground penetrating radar signals, and the system performs well in ballast fouling classification. Provided that the training data set is representative of antenna height variations, the system can operate with different antenna heights.

The second approach, motivated by the fact that GPR signals approximately resemble the Ricker wave (second-order derivative of Gaussian), decomposes each GPR trace into elementary waves using a dynamically expanding Gabor dictionary. The sparse decomposition is used to extract salient features for sparse representation and classification of GPR signals. It employs an over-complete Gabor dictionary that is dynamically refined during the sparse decomposition. Furthermore, the proposed adaptive signal decomposition is very effective for both signal representation and classification.

The third approach is based on time-frequency analysis. The frequency properties of GPR signals change with time. Hence, a time-frequency representation is useful to represent GPR signals. This approach utilizes the short time Fourier transformfor GPR signal representation and applies compressed sensing to select salient frequency components in the high-dimensional feature space as input features to a classifier. The experimental results show that the proposed approach performs well on real-world railway GPR data.

All the proposed approaches are evaluated on real-world data. The experimental results prove their efficiency in GPR signal representation and discriminative power for pattern classification using a small number of coefficients.