Target detection using supervised machine learning algorithms for GPR data

A novel approach of supervised machine learning technique is used in this paper to identify landmines. The work presented here has two contributions. First contribution is three features (major axis, minor axis and principal component analysis) based performance comparison of two machine learning technique: support vector machine classifier and neural network classifier. In the second contribution, a new method of extracting five features (mean, variance, kurtosis, skewness and entropy) is suggested. Support vector machine and neural network classifier are trained on three and five-feature data-set. Collection of ground penetrating radar images with surrogate landmines is done in our lab and a data-base of different feature set is created. In experiments, many surrogate mines and non-mines are considered at various depths for data collection. The performance of classifiers is compared on training and testing data-set. Out of the two classifiers, neural network classifier results with better accuracy of 85–90% for training data samples in both (three feature and five feature) analysis. Two trained classifiers are tested over twenty cases of unseen samples. Neural network classifier gives better results of 5–10% increased accuracy than support vector machine classifier over test set also.