基于深度学习的气象资料迹线识别

气象历史资料数字化是中国气象局“气候变化应对工程”的重要建设内容之一。气象历史资料数字化主要内容是完成珍贵纸质气象资料中字符和迹线信息的数字化处理,建立历史长序列气象资料数据集。气象自记纸记录的迹线信息是气象历史资料的重要内容,气象资料迹线信息的准确跟踪和提取是实现气象历史资料数字化的重要环节。然而,由于纸质气象资料存放时间长,气象迹线识别面临迹线记录模糊或迹线晕染等问题。传统的图像分割和迹线提取算法依赖局部或单一的图像特征,无法有效识别复杂背景中的气象迹线信息,特别是变化的、模糊的、晕染的迹线。为了解决上述问题,本文提出了一种基于深度学习模型的气象资料迹线自动识别方法。该方法结合传统算法和U-network (U-net)语义分割网络自动提取不同尺度的光谱和空间特征,实现气象迹线的高精度识别。本文以达因型风自记纸为例,以《地面气象观测规范》为依据,对不同台站、不同年份的3200多张达因风自记纸进行迹线识别验证,风向迹线平均识别正确率达95%,风速迹线平均识别正确率达95.5%。 结果表明该方法能够准确识别迹线,高度还原了气象迹线信息。该方法适用于风、气压、温度、湿度等气象资料迹线自动识别业务化处理,能够大大减少工作量和降低人工成本,为气象预报预测业务和科研工作提供基础数据支撑。

Recognition of meteorological curves based on deep learning

The digitization of meteorological historical data is one of the important construction contents of the "Climate Change Response Project" of the China Meteorological Administration. The main content of the digitization of meteorological historical data is to carry out the digital processing of characters and curves information in precious paper meteorological data, and to establish historical long-sequence meteorological data datasets. The curves information recorded by the meteorological self-recording paper is an important content of the meteorological historical data. The accurate tracking and extraction of the meteorological curves information is an important part of realizing the digitization of the meteorological historical data. However, due to the long storage time of paper meteorological data, the identification of meteorological curves has problems such as curves blurring or curves smearing. Traditional image segmentation and curves extraction algorithms rely on local or single image features, and cannot effectively identify meteorological curves information in complex backgrounds, especially changeable, fuzzy, and hazy curves. To solve the above problems, meteorological curves identification method based on deep learning was proposed in this paper. The method combines traditional algorithm and the U-net semantic segmentation network to automatically extract spectral and spatial features of different scales, and achieves high-precision meteorological curves recognition. This study takes dyne wind self-recording paper as an example, based on the "Ground Meteorological Observation Specification", more than 3200 dynes wind self-recording papers from different stations and in different years were verified. The comparison with manual reference results show that the average wind direction and wind speed recognition accuracy rate reached 95%, 95.5%, respectively. The results indicate that the method proposed in this paper can accurately identify the curves and highly restore the meteorological curves information. And the method is suitable for automatic identification and business processing of meteorological data curves such as wind, air pressure, temperature, and humidity, which can greatly reduce the workload and reduce labor costs, providing basic data support for meteorological forecasting business and scientific research.