气象科学2026，Vol.46Issue(1)：59-67,9.DOI:10.12306/2024jms.0046
基于改进卷积神经网络的雾天识别技术
Fog recognition technology based on improved CNN
摘要
Abstract
An improved convolutional neural network model for fog classification,based on convolutional neural network model fusion,incorporating a weighted attention mechanism,was introduced.Experimental evaluations utilizing fog image datasets from Lufeng,Xichou,Jiangcheng,and Honghe national weather stations in Yunnan Province,yield the following findings:(1)by incorporating a weighted attention mechanism,the enhanced convolutional neural network model demonstrates superior fog image feature extraction capabilities,achieving an average recognition accuracy of 95.69%.This surpasses the accuracy of traditional models like VGG and AlexNet.(2)When the trained model is applied to test fog image data from different time periods,it maintains an accuracy exceeding 90%at all four stations,with minimal identification errors among the test samples.This suggests that the model can serve as a valuable supplement to automatic weather station fog observations.(3)Analysis of erroneous test image cases reveals that the model's generalization ability may be compromised by the absence of certain training data.Additionally,rainfall can lead to misclassifications.Furthermore,the high similarity of image samples at the classification boundaries during foggy conditions can result in incorrect classifications by the model.关键词
注意力机制/卷积神经网络/雾天图像识别Key words
attention mechanism/convolutional neural network/fog image recognition分类
信息技术与安全科学引用本文复制引用
张国兴,张涛,马芳,和楷承,解莉燕..基于改进卷积神经网络的雾天识别技术[J].气象科学,2026,46(1):59-67,9.基金项目
云南省科技厅(202203AC100006-1) （202203AC100006-1）
中国气象局大气探测重点开放实验室开放课题(2023KLAS12M) （2023KLAS12M）
云南省气象局科研项目(YZ202305) （YZ202305）
