Outlier Detection Enhancement in Heterogeneous Environments Through a Novel Training Set Selection Framework

Most training set selection (TSS) methods are based on data processing methods. These methods have improved the state-of-the-art in clutter suppression under heterogeneous condition; however, TSS for heterogeneous and complex environments has rarely been investigated, especially for large outliers. This problem arises in situations such as isolated elevation points, spike effects of mountains, and urban-rural interfaces in actual radar operating environments. To address such a problem, this article proposes a novel enhanced outlier detection framework that deals with TSS in the presence of an unknown number of multiple outliers. First, the design of the overall structure of the TSS framework is proposed. We decompose the actual radar returns into four components and further integrate them into the TSS framework. The proposed framework uses the statistical characteristics of the returns from the range cells as a classification criterion. A deep neural network is devised to extract these statistical characteristics for outlier detection. The loss function and learning rate selection of the proposed TSS framework are, furthermore, specified. Then, the classification model for the four signal components is presented. To validate this framework, we use a real radar dataset sampled from heterogeneous environments and characterize signals in real radar scenarios. Experimental results demonstrate that the proposed framework significantly improves the accuracy of outlier detection in comparison with the traditional heterogeneous TSS method. In addition, our framework can further distinguish the interference outliers from the target echoes.