基于雷达目标特征可分性的一维特征选择方法

One-Dimensional Feature Selection Method Based on Radar Target Feature Divisibility

海杂波背景下的海上小目标是海洋雷达探测的重难点。针对特征空间内海杂波与小目标特征可分性问题,提出了量化特征之间可分性的度量标准——重叠系数。通过开展对海探测试验获取的2~5级海况实测数据,分别提取时域特征相对平均幅度(Relative Average Amplitude,RAA)、相对峰值峰高(Relative Peak Height,RPH)、时域熵值均值(Time domain Entropy Mean,TEM),频域特征相对多普勒峰高(Relative Doppler Peak Height,RDPH)、相对多普勒向量熵(Relative Vector Entropy,RVE)、频域熵值二阶矩(Second moment of Frequency domain Entropy,SOFE),计算出重叠系数。通过特征检测器进行检测性能对比,低海况下,相对平均幅度、相对峰值峰高、时域熵值均值、相对多普勒峰高、频域熵值二阶矩特征之间重叠系数均在0.3以下,对应特征检测器的检测概率均在85%以上;高海况下其特征之间重叠系数均在0.7以上,对应特征检测器的检测概率均在50%以下。相对多普勒向量熵在4种海况下可分性较小,其对应的特征检测器性能较差。结果验证了重叠系数在一维特征选择的应用可行性,为多特征融合目标检测提供了一定支持。

Small targets on the sea under the background of sea clutter are the key and difficult points of ocean radar detection.It is proposed that a measurement standard for quantifying the separability between features-overlap coefficient to response to the issue of separability between sea clutter and small target features in the feature space.By conducting sea detection experiments on measured sea conditions at levels 2~5,the relative average amplitude,relative peak height,and mean time domain entropy of time-domain features are extracted,as well as the relative Doppler peak height,relative Doppler vector entropy,and second-order moment of frequency-domain entropy of frequency-domain features,the overlap coefficient is calculated.By comparing the detection performance through feature detectors,under low sea conditions,the overlap coefficients between the relative average amplitude,relative peak height,time-domain entropy mean,relative Doppler peak height,and frequency-domain entropy second-order moment features are all below 0.3,and the detection probability of the corresponding feature detectors is above 85%;under high sea conditions,the overlap coefficients between its features are all above 0.7,and the detection probability of the corresponding feature detectors is below 50%.Relative vector entropy has low separability under four sea conditions,and its corresponding feature detectors have poor performance.The conclusion verifies the feasibility of applying overlap coefficients in one-dimensional feature selection,providing some support for multi-feature fusion target detection.