摘要
精确识别冰相粒子对云降水机制理解、天气作业条件判断及降水预测至关重要。本文提出大核筛选注意力网络:以多尺度卷积耦合空间筛选萃取判别特征,非对称卷积削减冗余参数;广义平均池化全卷积分类器强化变换粒子鲁棒表征。实验表明,相较VAN-b1,准确率、精确率、召回率分别提升3.02%、3.26%、4.26%,参数量降低28.98%,实现高精度冰相粒子识别,为云降水机制研究与作业条件判别提供技术支撑。
Accurate identification of ice-phase particles is crucial for understanding cloud-prccipi-tation mechanisms,assessing weather operation conditions,and improving precipitation forecas-ting.This paper proposes a large-kernel screening attention network:multi-scale convolutional coupling space screening extracts discriminative features,while asymmetric convolutions reduce redundant parameters;generalized average pooling and fully convolutional classifiers enhance robust characterization of transformed particles.Experiments demonstrate that compared to VAN-bl,accuracy,precision,and recall improve by 3.02%,3.26%,and 4.26%respectivc-ly,while parameters decrease by 28.98%.This achieves high-precision ice particle identifica-tion,providing technical support for cloud-precipitation mechanism research and operational condition assessment.
作者
钟小伟
胡金蓉
刘星光
ZHONG Xiaowei;HU Jinrong;LIU Xingguang(school of Computer Science,Chengdu University of Information Science Technology,Chengdu,Sichuan 610200,China;Key Laboratory of Smart Earth,Beijing 100029,China;innovation and Opening Laboratory of Eco-Meteorology in Northeast China,CMA,Harbin,Heilongjiang 150030,China;Heilongjiang Province Institute of Meteorological Sciences,Heilongjiang 150030,China)
出处
《长江信息通信》
2025年第8期62-65,共4页
Changjiang Information & Communications
基金
中国气象局创新发展专项(CXFZ2022J038,CXFZ2024J035)
四川省科技计划项目(2023YFQ0072)
智慧地球重点实验室基金资助项目(KF2023YB03-07)。
关键词
冰相粒子分类
大核注意力
多尺度
机载探测
Ice crystal classification
Large kernel attention mechanism
Multi-scale
Airborne detection
