快速高效的张衡一号卫星ELF波段闪电哨声波自动检测算法  

A fast and efficient algorithm for automatic detection of ELF lightning whistles recorded by the ZH-1 satellite

作　　者：赵晨旭 袁静[1] 王桥 申旭辉 泽仁志玛 刘庆杰[1] 黄建平 刘祖阳 刘海军[1] ZHAO ChenXu;YUAN Jing;WANG Qiao;SHEN XuHui;Zeren ZhiMa;LIU QingJie;HUANG JianPing;LIU ZuYang;LIU HaiJun(Institute of Disaster Prevention,Langfang Hebei 065201,China;Institute of Geophysics,China Earthquake Administration,Beijing 100081,China;National Space Science Center,Chinese Academy of Sciences,Beijing 100190,China;National Institute of Natural Hazards,Ministry of Emergency Management of China,Beijing 100085,China;Hubei University,Wuhan 430062,China)

机构地区：[1]防灾科技学院,河北廊坊065201 [2]中国地震局地球物理研究所,北京100081 [3]中国科学院国家空间科学中心,北京100190 [4]应急管理部国家自然灾害防治研究院,北京100085 [5]湖北大学,武汉430062

出　　处：《地球物理学报》2024年第11期4089-4104,共16页Chinese Journal of Geophysics

基　　金：河北省教育厅科学研究项目(ZC2024028);国家自然科学基金青年基金(42104159)资助。

摘　　要：张衡一号卫星在轨六年积累了大量的闪电哨声波(Lightning Whistlers,LWs)事件,这些事件是全面研究空间物理环境和圈层耦合机理的重要媒介.依靠目前的智能算法完成LWs识别任务需要几十年的时间,难以满足实际工程需要.本文提出了一种快速高效的闪电哨声波自动识别算法(Light Weight Network for Lightning Whistler,LW-LWNet):首先,采用深度可分离卷积、挤压激励机制等轻量化技术对YOLOv5目标检测算法的主干网络进行改进,通过降低参数量和计算量,提高了模型推理速度;其次,采用小计算量的注意力机制改进主干网络的输出通道,通过增强闪电哨声波形态特征,克服因参数压缩导致性能下降的问题;最后,通过训练得到本文提出的LW-LWNet模型.在2019年9月的LWs数据集上进行的实验表明,LW-LWNet模型在精确度、召回率、准确率和F1分别达到了88.8%、80.6%、89.8%和89.3%,相对于原始算法提高了0.7%、0.9%、0.4%和0.6%.此外,在轻量化方面,该模型的参数量减少了57%;在推理速度方面,FPS提升33%;在检测精度方面,mAP50提升了0.3%.ZH-1 satellite has accumulated substantial data of Lightning Whistlers (LWs) over its six years in orbit, serving as crucial tools for comprehensive study of the space physical environment and inter-layer coupling mechanisms. However, the current algorithms require decades to identify LWs, which is impractical for engineering applications. To address this, we propose a fast and efficient Lightweight Network(LW-LWNet)for detecting lightning whistlers. Our approach utilizes lightweight technologies such as depth-separable convolution and squeeze excitation mechanism to enhance the backbone network of YOLOv5 target detection algorithm. This reduces parameters and computational complexity, thereby improving inference speed. Additionally, we employ a small computational attention mechanism to improve the backbone network's output channels, highlighting the characteristics of lightning whistle waves and mitigating performance degradation due to parameter compression. The LW-LWNet model was trained and evaluated on LW datasets from September 2019, achieving an accuracy of 88.8%, a recall of 80.6%, a precision of 89.8%, and an F1 score 89.3%. These results represent improvements of 0.7%, 0.9%, 0.4%, and 0.6% respectively over the original algorithm. Furthermore, the model's parameters were reduced by 57%, inference speed (FPS) increased by 33%, and detection accuracy (mAP50) improved by 0.3%. Experiments demonstrate that the LW-LWNet model not only enhances recognition accuracy but also significantly boosts inference speed, offering an effective reference for exploring the temporal and spatial distribution of global lightning whistlers.

关 键 词：闪电哨声波 YOLOv5 轻量化技术 注意力机制 

分 类 号：P353[天文地球—空间物理学]