Abstract:

Recognition of weather phenomena based on images is essential for the analysis of weather conditions. To address the problems that traditional machine learning methods are difficult to accurately extract various weather features and poor in classifying weather phenomena and the accuracy of deep learning for weather phenomena recognition is not high, a weather recognition model based on image block and multi?headed attention mechanism is proposed. The model introduces Swin Transformer into the field of weather recognition for the first time, and adopts a multi?headed attention mechanism combining window multi?head self?attention layer and shifted?window multi?head self?attention layer, whose regionally relevant features extraction capability makes up for the shortcomings of traditional methods and can extract complex weather features from images. The model is trained using transfer learning, and the fully connected parameters of the fine?tuned model are input to the Softmax classifier to achieve recognition of multi?category weather images with 99.20% recognition accuracy, which is better than several mainstream methods in comparison, and it can be applied to ground weather recognition systems as a weather recognition module.

Key words: Weather recognition, Image block, Multi?headed attention mechanism, Regional correlation features, Transfer learning

摘要：

基于图像对天气现象进行识别，对天气状况的分析至关重要。针对传统的机器学习方法对各类天气特征难以准确提取且天气现象分类效果差，以及深度学习对天气现象识别的准确率不高的问题，提出了基于图像分块和多头注意力机制的天气识别模型。该模型首次将Swin Transformer引入天气识别领域中，采用了窗口多头自注意层与移位窗口多头自注意层相结合的多头注意力机制。结果表明，其区域相关特征提取能力弥补传统方法的不足，能够提取图像中复杂的天气特征。采用迁移学习对模型进行训练，将微调模型的全连接参数输入到Softmax分类器，实现了对多类别天气图像的识别，识别准确率为99.20%，优于对比的几种主流方法。因此，该方法可以作为天气识别模块应用于地面气象识别系统。

关键词: 天气识别, 图像分块, 多头注意力机制, 区域相关特征, 迁移学习