阵风锋是一种常见的中小尺度天气现象,其经过常伴有灾害性天气,对生命财产安全构成严重威胁,因此,精确检测与识别阵风锋对于防灾减灾具有重要的现实意义。然而在实际应用中,阵风锋的检测面临着数据稀缺、区域特异性强、传统方法检测效...阵风锋是一种常见的中小尺度天气现象,其经过常伴有灾害性天气,对生命财产安全构成严重威胁,因此,精确检测与识别阵风锋对于防灾减灾具有重要的现实意义。然而在实际应用中,阵风锋的检测面临着数据稀缺、区域特异性强、传统方法检测效果不佳、形态与位置难以精确识别、以及泛化能力不足等挑战。此外阵风锋的特征不明显,易与其他大气现象混淆,导致高误判率的发生。针对这些问题,本文提出了一种基于深度学习的创新方法。首先,针对传统检测方法的局限性,本文改进了Mask R-CNN模型,并引入了注意力机制和特征融合模块,显著提升了阵风锋的检测精度。其次,本文通过引入径向速度频道信息并设计上下文分支,有效增强了模型对干扰要素的辨别能力,从而降低了误判率。此外,本文构建了基于新一代多普勒天气雷达数据的三维阵风锋数据集,并通过数据增强技术扩充样本量,为阵风锋的深入研究提供了充足的资源。实验结果表明,所提出的方法在提升阵风锋检测与识别准确性方面具有显著优势,为阵风锋的研究及其业务化应用提供了新的理论基础和技术支持。Gust fronts are common mesoscale meteorological phenomena that are frequently accompanied by severe weather events, posing significant threats to life and property. Accurate detection and identification of gust fronts are therefore of critical importance for disaster prevention and mitigation. However, several challenges persist in the detection of gust fronts, including data scarcity, strong regional specificity, suboptimal performance of traditional mathematical methods, difficulties in accurately identifying their shape and location, and limited generalization capability. Moreover, the subtle characteristics of gust fronts often lead to confusion with other atmospheric phenomena, resulting in high false positive rates. To address these challenges, this paper proposes an innovative approach based on deep learning techniques. First, to overcome the limitations of traditional detection methods, we improve the Mask R-CNN model by incorporating an attention mechanism and a feature fusion module, significantly enhancing the detection accuracy of gust fronts. Second, to reduce false positive rates, we introduce radial velocity channel information and design a context branch to strengthen the model’s ability to distinguish between gust fronts and interfering elements. Additionally, we develop a three-dimensional gust front dataset using next-generation Doppler weather radar data and expand the dataset through data augmentation techniques, thereby providing a robust resource for gust front research. Experimental results validate the effectiveness of the proposed method in enhancing the accuracy of gust front detection and identification, offering new perspectives and tools for both research and operational applications in this domain.展开更多
0~2小时的临近预报是预报强对流天气的重要手段,准确的临近预报是气象防灾的重要屏障。临近预报是气象预报的一个重要业务。本文提出了一个基于时序Transformer的全局特征提取的Encoder-Forecasting临近预报模型结构。该模型的Transfor...0~2小时的临近预报是预报强对流天气的重要手段,准确的临近预报是气象防灾的重要屏障。临近预报是气象预报的一个重要业务。本文提出了一个基于时序Transformer的全局特征提取的Encoder-Forecasting临近预报模型结构。该模型的Transformer利用双级路由注意力机制定位特征中的几个最相关的键值对来提高计算效率。模型同时采用MaxPool和AvgPool提取目标对象的局部特征,并将局部特征和全局特征进行有效融合,进而充分提取目标对象的特征用于临近预报。提出的模型与几个模型在一个公共的临近预报数据集上进行实验验证。实验结果充分证明了我们提出的模型的有效性和准确性。The nowcast of 0~2 hour is an important toolth for forecasting severe convective weather. Accurate nowcast is an important barrier to prevent from weather disaster. Therefore, nowcast is a very important field in the meteorological society. This paper proposes a nowcast model which based on Encoder-Forecasting architecture and temporal Transformer. The transformer utilizes bilevel routing attention to improve the computational efficiency. At the same time, the proposed model also use the operations of Maxpool and AvgPool to extract the local features. The global features and local features are effectively fused to represent the objects’ features. The comprehensive experiments based on the proposed model and four state-of-art models are conducted on a public datasets. The experimental results effectively show the correctness and effectiveness of the proposed model.展开更多
文摘阵风锋是一种常见的中小尺度天气现象,其经过常伴有灾害性天气,对生命财产安全构成严重威胁,因此,精确检测与识别阵风锋对于防灾减灾具有重要的现实意义。然而在实际应用中,阵风锋的检测面临着数据稀缺、区域特异性强、传统方法检测效果不佳、形态与位置难以精确识别、以及泛化能力不足等挑战。此外阵风锋的特征不明显,易与其他大气现象混淆,导致高误判率的发生。针对这些问题,本文提出了一种基于深度学习的创新方法。首先,针对传统检测方法的局限性,本文改进了Mask R-CNN模型,并引入了注意力机制和特征融合模块,显著提升了阵风锋的检测精度。其次,本文通过引入径向速度频道信息并设计上下文分支,有效增强了模型对干扰要素的辨别能力,从而降低了误判率。此外,本文构建了基于新一代多普勒天气雷达数据的三维阵风锋数据集,并通过数据增强技术扩充样本量,为阵风锋的深入研究提供了充足的资源。实验结果表明,所提出的方法在提升阵风锋检测与识别准确性方面具有显著优势,为阵风锋的研究及其业务化应用提供了新的理论基础和技术支持。Gust fronts are common mesoscale meteorological phenomena that are frequently accompanied by severe weather events, posing significant threats to life and property. Accurate detection and identification of gust fronts are therefore of critical importance for disaster prevention and mitigation. However, several challenges persist in the detection of gust fronts, including data scarcity, strong regional specificity, suboptimal performance of traditional mathematical methods, difficulties in accurately identifying their shape and location, and limited generalization capability. Moreover, the subtle characteristics of gust fronts often lead to confusion with other atmospheric phenomena, resulting in high false positive rates. To address these challenges, this paper proposes an innovative approach based on deep learning techniques. First, to overcome the limitations of traditional detection methods, we improve the Mask R-CNN model by incorporating an attention mechanism and a feature fusion module, significantly enhancing the detection accuracy of gust fronts. Second, to reduce false positive rates, we introduce radial velocity channel information and design a context branch to strengthen the model’s ability to distinguish between gust fronts and interfering elements. Additionally, we develop a three-dimensional gust front dataset using next-generation Doppler weather radar data and expand the dataset through data augmentation techniques, thereby providing a robust resource for gust front research. Experimental results validate the effectiveness of the proposed method in enhancing the accuracy of gust front detection and identification, offering new perspectives and tools for both research and operational applications in this domain.
文摘0~2小时的临近预报是预报强对流天气的重要手段,准确的临近预报是气象防灾的重要屏障。临近预报是气象预报的一个重要业务。本文提出了一个基于时序Transformer的全局特征提取的Encoder-Forecasting临近预报模型结构。该模型的Transformer利用双级路由注意力机制定位特征中的几个最相关的键值对来提高计算效率。模型同时采用MaxPool和AvgPool提取目标对象的局部特征,并将局部特征和全局特征进行有效融合,进而充分提取目标对象的特征用于临近预报。提出的模型与几个模型在一个公共的临近预报数据集上进行实验验证。实验结果充分证明了我们提出的模型的有效性和准确性。The nowcast of 0~2 hour is an important toolth for forecasting severe convective weather. Accurate nowcast is an important barrier to prevent from weather disaster. Therefore, nowcast is a very important field in the meteorological society. This paper proposes a nowcast model which based on Encoder-Forecasting architecture and temporal Transformer. The transformer utilizes bilevel routing attention to improve the computational efficiency. At the same time, the proposed model also use the operations of Maxpool and AvgPool to extract the local features. The global features and local features are effectively fused to represent the objects’ features. The comprehensive experiments based on the proposed model and four state-of-art models are conducted on a public datasets. The experimental results effectively show the correctness and effectiveness of the proposed model.
