融合LightGBM的ResNeXt气象目标细粒度识别方法  

Fine-grained recognition method for meteorological targets in ResNeXt fused with LightGBM

作　　者：欧阳彤 汪玲[1] 朱岱寅[1] 李勇[1] OUYANG Tong;WANG Ling;ZHU Daiyin;LI Yong(College of Electronic and Information Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学电子信息工程学院,江苏南京211106

出　　处：《系统工程与电子技术》2024年第12期4034-4043,共10页Systems Engineering and Electronics

基　　金：工信部民机专项(MJ-2018-S-28)资助课题。

摘　　要：为精确识别气象目标与混杂其中的非气象目标,提出一种融合轻量级梯度提升机(light gradient boosting machine,LightGBM)与残差网络的残差网络(residual network of residual network:next generation,ResNeXt)的气象目标识别方法。首先,制作块状样本数据集,以此数据集为驱动,建立以ResNeXt为基础的气象目标识别网络模型,实现以块状数据样本为识别单位的气象目标粗粒度识别,识别精度可达99.6%以上;然后,再将此粗粒度结果与参考数据的差异值纳入LightGBM分类器,得到以雷达采样单元为识别单位的细粒度识别结果。结合实际观测数据,证明所提方法融合了LightGBM细粒度识别与ResNeXt高精度识别的能力,能够完成气象目标与杂波的判别,判别结果与参考结果高度一致。结合实际观测数据,证明所提方法融合了LightGBM细粒度识别与ResNeXt高精度识别的能力,能够完成气象目标与杂波的判别,判别结果与参考结果高度一致。In order to accurately identify meteorological targets and non-meteorological targets mixed in the meteorological targets,a meteorological target recognition method that combines light gradient boosting machine(LightGBM)and residual network of residual network:next generation(ResNeXt)network is proposed.Firstly,a block sample dataset is produced,and this dataset is used as a driver to establish a meteorological target recognition network model based on the ResNeXt.This model realizes a coarse-grained recognition of meteorological targets with block data samples as the recognition unit with the recognition accuracy of more than 99.6%.Then,the difference between the coarse-grained recognition results and the reference data is included in the LightGBM classifier so the fine-grained recognition results with the radar sampling unit as recognition unit are obtained.Combined with the actual observation data,it is proved that the proposed method combines the ability of fine-grained recognition of LightGBM with the ability of high-precision recognition of ResNeXt,which can distinguish meteorological targets from clutter,and the distinguishing results are highly consistent with the reference results.

关 键 词：气象雷达 气象目标识别 残差网络 轻量级梯度提升机 融合 深度学习 

分 类 号：TN957.5[电子电信—信号与信息处理]