Classification of GPR data for mine detection based on hidden Markov models

We present a novel approach for the classification of GPR data, based on hidden Markov models. It assumes that the system, generating the recorded data, can be in one of a set of distinct states. At discrete intervals, given by the distance between the recording positions of two adjacent radar scans, the system can either undergo a change of state or remain in the same state, according to a set of probabilities assigned to the allowed transitions between states. The appeal of the method is that it is not restricted to a classification on a scan-by-scan basis, but that it allows one to look at a sequence of data of a certain lateral extension. This approach can thus accommodate characteristic object pattern evolving not only in time, but also in space. Our results indicate that HMMs outperform scan-wise classification, based on alternative algorithms, such as polynomial classifiers, neural or radial basis function networks.