Feature extraction of ground penetrating radar for mine detection

Feature extraction is applied to mine detection data from a downward-looking ground penetrating radar(GPR) array. GPR signals have low signal-to-clutter ratio, are non-stationary in space, and vary with humidity, temperature, and soil moisture. To enhance mine-like signals and suppress false alarms, overlapping sensors allow one dimensional sensor fusion and adjacent sensors allow two dimensional sensor fusion. Maximum likelihood estimation followed by template matching perform confident detections in discriminating suspicious locations. The algorithm includes a training phase and a testing phase. In the training phase, local clutter features and their largest ten eigenvectors are extracted from known clean data using principal component analysis. In the testing phase, local background clutter is first removed from the raw data using a moving-average filter. Secondly, the de-cluttered data is projected on the significant clutter eigenvectors developed in training phase. A binary decision is made at each pixel according to template matching distances and geometric sensor structure. Receiver operating curve evaluations against test bed ground truth show improvement as singular value decomposition is enhanced by template matching and 1-D and 2-D sensor fusion.