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。
Fine-grained recognition method for meteorological targets in ResNeXt fused with LightGBM
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.
万方数据
维普
