Sea-surface Floating Small Target Detector Combining Decision Tree with Anomaly Detection in High-Dimensional Feature Space

Due to the complex and diverse characteristics of sea clutter and small sea-surface targets in high-resolution radar echoes, there is currently no strict and simple statistical model that can describe the distributions of sea clutter and targets. Instead, multiple statistics (termed features) without strict models are employed for detection. In a high-dimensional (HD) feature space, the problem of sea-surface small target detection can be transformed into a binary hypothesis classification problem, which is further converted into classifier design with two unbalances. There exists an imbalance between sufficient and ergodic sea clutter samples and scarce and non-ergodic target-containing samples. The false alarm rate must not exceed 10³, while a miss probability of several tenths is permissible. To obtain adequate training samples for the alternative hypothesis, a target echo generator is developed to simulate target echoes and generate training samples. Considering the non-ergodic characteristic of target echoes, a decision tree-oriented detector featuring a controllable false alarm rate is used as the classifier. Through pre-decision by anomaly detection (based on sea clutter characteristics), unclassified test samples undergo further decision-making via the decision tree, resulting in a new detection system. Experiment results from the acknowledged and publicly available IPIX and CSIR radar databases, in comparison to current feature-oriented detectors, verify that the proposed detection mechanism brings about notable performance enhancements.