Sea-Surface Floating Small Target Detection by Multifeature Detector Based on Isolation Forest

In this article, a multifeature detector based on isolation forest (iForest) algorithm is developed to detect floating small targets in sea clutter. The conventional multifeature detector can only process three features or less. The proposed detector aims to break the limitation of feature dimensions' number of the existed feature-based detectors and to improve the detection performance. It transforms the detection of floating small target into an anomaly detection problem in a high-dimensional feature space, breaking the limitation of the number of features. First, a modified isolation forest is constructed from multiple features extracted from sea clutter. Meanwhile, the relative Doppler coefficient of variation is proposed and added into the feature library. Then, taking the average path length as detection statistic, the detection threshold is obtained by Monte-Carlo technique at the given false alarm probability. Finally, the final decision is made by comparing the path length calculated from the cell under test of radar returns with the detection threshold. Detection performances are evaluated based on twenty measured IPIX radar datasets. The experiment results show that the multifeature detector based on isolation forest can obtain a significant performance improvement and has lower computation cost compared with the existed detectors.