当前位置: X-MOL 学术Atmos. Res. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
An automatic trough line identification method based on improved UNet
Atmospheric Research ( IF 4.5 ) Pub Date : 2021-08-30 , DOI: 10.1016/j.atmosres.2021.105839
Yali Cai 1 , Qian Li 1 , Yin Fan 1 , Liang Zhang 1 , Hong Huang 1 , Xinya Ding 1
Affiliation  

A trough is an elongated region of relatively low atmospheric pressure. Automatic analysis and recognition of trough lines in upper weather charts are challenging and of great significance to weather analysis. The existing automatic identification methods mainly depend on manually setting rules which cannot cover all trough line types and have low generalization ability. This paper proposes an automatic trough line identification method based on an improved model which can extract the trough line from meteorological element data at 500 hPa. The model adopts the UNet of a U-shaped encoder and decoder as basic structure, which is designed to enable precise localization by continuously combining low-level and high-level features. To extract abstract semantic features of the trough, the Xception, which takes depthwise separable convolution as basic unit, is adopted to replace the encoder of the original UNet. In addition, the Squeeze and Excitation (SE) module with an attention mechanism is added after each ordinary convolution in the decoder part to improve the recognition accuracy by increasing the weighting of the trough area. The experiments are conducted on a meteorological dataset and the results show that the recognition accuracy with our proposed method on the testing dataset can reach over 80%. We also compare our results to several other types of networks and traditional automatic identification methods, which demonstrates that the performance of the proposed network is superior to other methods.



中文翻译:

一种基于改进UNet的槽线自动识别方法

槽是大气压力相对较低的细长区域。上层天气图中槽线的自动分析和识别具有挑战性,对天气分析具有重要意义。现有的自动识别方法主要依靠人工设置规则,不能覆盖所有槽线类型,泛化能力低。本文提出了一种基于改进模型的自动槽线识别方法,该方法可以从500 hPa的气象要素数据中提取槽线。该模型采用U形编码器和解码器的UNet作为基本结构,旨在通过不断结合低层和高层特征来实现精确定位。为了提取槽的抽象语义特征,Xception,采用以深度可分离卷积为基本单元的,取代原来的UNet的编码器。此外,解码器部分在每次普通卷积后都加入了带有attention机制的Squeeze and Excitation (SE)模块,通过增加波谷区域的权重来提高识别精度。在气象数据集上进行了实验,结果表明,我们提出的方法在测试数据集上的识别准确率可以达到80%以上。我们还将我们的结果与其他几种类型的网络和传统的自动识别方法进行了比较,这表明所提出的网络的性能优于其他方法。在解码器部分的每个普通卷积之后添加具有注意力机制的Squeeze and Excitation (SE)模块,通过增加波谷区域的权重来提高识别精度。在气象数据集上进行了实验,结果表明,我们提出的方法在测试数据集上的识别准确率可以达到80%以上。我们还将我们的结果与其他几种类型的网络和传统的自动识别方法进行了比较,这表明所提出的网络的性能优于其他方法。在解码器部分的每个普通卷积之后添加具有注意力机制的Squeeze and Excitation (SE)模块,通过增加波谷区域的权重来提高识别精度。在气象数据集上进行了实验,结果表明,我们提出的方法在测试数据集上的识别准确率可以达到80%以上。我们还将我们的结果与其他几种类型的网络和传统的自动识别方法进行了比较,这表明所提出的网络的性能优于其他方法。在气象数据集上进行了实验,结果表明,我们提出的方法在测试数据集上的识别准确率可以达到80%以上。我们还将我们的结果与其他几种类型的网络和传统的自动识别方法进行了比较,这表明所提出的网络的性能优于其他方法。在气象数据集上进行了实验,结果表明,我们提出的方法在测试数据集上的识别准确率可以达到80%以上。我们还将我们的结果与其他几种类型的网络和传统的自动识别方法进行了比较,这表明所提出的网络的性能优于其他方法。

更新日期:2021-09-02
down
wechat
bug