基于代数多重网格的天波超视距雷达跨尺度地海杂波识别方法

Cross-Scale Land/Sea Clutter Classification Method for Over-the-Horizon Radar Based on Algebraic Multigrid

天波超视距雷达(天波雷达)在远程预警领域发挥着关键作用.基于天波雷达地海杂波识别的坐标配准利用地海杂波识别结果形成地/海分界线或地形轮廓,将其与先验地理信息匹配为目标定位提供坐标配准参数,可提升天波雷达目标定位精度.为满足不同类型目标检测、波束驻留与扫描等要求,天波雷达通常采用不同信号时宽、相干积累点数,使地海杂波谱数据具有多分辨率多尺度特性.针对不同分辨率/尺度地海杂波谱数据分别设计分类器存在训练数据不均衡、维护成本高等问题.本文基于代数多重网格与插值相关图像下采样思想,建立不同尺度地海杂波谱数据之间的代数关系,提出了一种跨尺度深度卷积神经网络地海杂波分类器.其允许使用经过训练的低分辨率地海杂波分类器对高分辨率数据进行分类,分类正确率不低于88.26%;也允许使用经过训练的高分辨率地海杂波分类器对低分辨率数据进行分类,分类正确率不低于92.53%,而无需针对不同分辨率/尺度数据分别设计分类器.

Skywave over-the-horizon radar(OTHR) plays a key role in early warning of long-range targets. Coordinate registration based on OTHR land/sea clutter recognition uses the land/sea clutter recognition results to construct the land/sea boundary or terrain contour, which is matched with the a priori geographic information to provide coordinate registration parameters for target positioning, which can improve the target positioning accuracy of OTHR. To meet the requirements of different types of target detection, OTHR usually adopts different signal time-widths and coherent integration points, so that the land/sea clutter spectorum data has multi-resolution and multi-scale characteristics. Designing classifiers for different resolution/scale land/sea clutter spectrum data has the problems of unbalanced training data and high maintenance costs. Based on the idea of algebraic multigrid and interpolation-dependent image downsampling, this paper establishes the algebraic relationship between the land/sea clutter spectrum data of different scales and proposes a cross-scale deep convolution neural network land/sea clutter classifier. This method allows us to use trained low-resolution land/sea clutter classifiers to classify high-resolution data. The classification accuracy is no less than 88.26%. It also allows us to use a trained high-resolution land-sea clutter classifier to classify low-resolution data. The classification accuracy is no less than 92.53%. It is not necessary to separately design classifiers for spectrum with different scales.