基于深度学习的全天空相机成像日间云量计算研究  

Research on Daytime Cloudiness Calculation for All-sky Camera Imagery Based on Deep Learning

作　　者：车蕾[1] 李磊磊 刘立勇[2] CHE Lei;LI Lei-lei;LIU Li-yong(School of Information Management,Beijing Information Science and Technology University,Beijing 100192,China;National Astronomical Observatories,Chinese Academy of Sciences,Beijing 100101,China)

机构地区：[1]北京信息科技大学信息管理学院,北京100192 [2]中国科学院国家天文台,北京100101

出　　处：《天文学进展》2024年第2期349-361,共13页Progress In Astronomy

基　　金：国家重点研发计划(YS2021YFC2203202);国家自然科学基金(11873063);全国高等院校计算机基础教育研究会计算机基础教育教学研究课题(2023-AFCEC-004)。

摘　　要：云量是天文领域中地基光电望远镜站址选择的重要评价参数之一。针对全天空相机成像的日间云量计算存在的问题,提出一种基于深度学习的全天空相机成像日间云量计算模型。云检测层,模型通过构建通道加权-特征融合(channel weighting-feature fusion,CWFF)结构,从而加强对云层记忆能力和深层特征的提取能力以完成云检测任务;云量计算层,模型提出一种基于云检测模型的云量计算方法,有效提高云量计算的误差率。实验表明,该方法在云检测任务中的综合准确率超过95%,在云量计算任务中的平均绝对误差不超过5%。Cloudiness is one of the important evaluation parameters for the site selection of ground-based photoelectric telescopes in astronomical field.The traditional cloudiness calculation method has a large deviation in the accuracy of cloudiness calculation for all-sky camera imagery,which is difficult to meet the actual demand for the accuracy of cloudiness calculation in multiple fields,and there are some limitations in its detection model extraction capability.Aiming at the problems of daytime cloudiness calculation of all-sky camera imaging,a deep learning-based daytime cloudiness calculation model of all-sky camera imaging is proposed.In the cloudiness detection layer,the model constructs a Channel Weighting-Feature Fusion(CWFF)structure to enhance the cloud memory and deep feature extraction capability to accomplish the cloudiness detection task.In the cloudiness calculation layer,the model proposes a cloudiness calculation method based on the cloudiness detection model,which effectively improves the error rate of cloudiness calculation.Experiments show that the combined accuracy of this paper’s method in the cloudiness detection task exceeds 95%,and the average absolute error in the cloudiness volume calculation task does not exceed 5%.

关 键 词：全天空相机 云量计算 深度学习 U型网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]