Multispectral target detection by statistical methods

In this study, targets and nontargets in a multispectral image were characterized in terms of their spectral features. Then, target detection procedures were performed. Target detection problem was considered as a two-class classification problem with four-band (Red-Green-Blue-Near Infrared) images. For this purpose, statistical techniques were employed. These are Parallelepiped, Euclidean Distance and Maximum Likelihood (ML) algorithms, which belong to supervised statistical classification methods. To obtain the training data belonging to each class, the training regions were selected as polygonal. After determination of the parameters of the algorithms with the training set, classification was accomplished at each pixel as target or background. Consequently, classification results were displayed on thematic maps. The algorithms were trained with the same training sets, and their comparative performances were tested under various situations. During these studies, the effects of training area selection and various levels of thresholds were evaluated based on the efficiency of the algorithms. The selection of appropriate technique was proposed, dependent upon different kinds of targets. The training area selection especially affected the performance of the ML algorithm. In spite of the fact that the training area selected as a target class did not vary, insufficient representation of the background classes in terms of training area resulted in high false alarm rate. Good representation of the background classes in the training set increased the detection rate while the false alarm rate was very much decreased. The training area selection was less critical with the performances of the Euclidean Distance and the Parallelepiped algorithms. These were more heavily dependent on the target training area.