面向气象观测的全天空云状智能检测技术研究

Research on Intelligent All-Sky Cloud Pattern Recognition Technology for Meteorological ObservationZHAO Shang-zhuo12, ZHANG Zi-yi13, CUI Juan*14

摘要:

为提升气象观测中云状检测的准确性，提出一种基于改进深度学习的全天空云状智能检测技术。在传统深度学习中引入多尺度特征金字塔融合模块，自适应捕捉云层从细微纹理到宏观结构的多尺度特征；同时添加注意力模块，动态加权关键特征区域，有效抑制不重要特征的干扰。输入全天空云状图像到基于改进深度学习的智能检测模型当中，计算每种云状类别的分布概率，取最大值对应的结果作为检测结果。基于华北地区的全天空观测数据，构建了包含10类典型云状的数据集，进行实验测试。实验结果表明:所提出的检测技术在消融实验中的交叉熵损失从0.0088逐渐下降到0.0004；对比实验中，Matthews相关系数稳定在0.91~0.98之间，性能显著更优，检测准确性更高。

Abstract:

To enhance the accuracy of cloud pattern recognition in meteorological observation,an intelligent all-sky cloud pattern recognition technology based on improved deep learning is proposed.By introducing a multi-scale feature pyramid fusion module into traditional deep learning,it adaptively captures cross-scale features of clouds from fine textures to macroscopic structures.Meanwhile,an attention module is added to dynamically weight key feature regions and effectively suppress the interference of unimportant features.The all-sky cloud pattern images are input into the intelligent recognition model based on improved deep learning to calculate the distribution probability of each cloud pattern category,and the result corresponding to the maximum value is taken as the recognition result.Based on the all-sky observation data in North China,a dataset containing 10 typical cloud patterns was constructed for experimental testing.The experimental results show that the cross-entropy loss in the ablation experiment gradually decreases from 0.0088 to 0.0004;in the comparison experiment,the Matthews correlation coefficient is stably between 0.91 and 0.98,demonstrating significantly better performance and higher recognition accuracy.