SAR-UNet:一种基于空间注意力机制的大气河识别网络模型

SAR-UNet: A Model of Atmospheric River Recognition Network Based on Spatial Attention Mechanism

准确、高效地识别大气河可以有效地预防洪涝和干旱等自然灾害,对地区的经济、社会和生态发展具有非常重大的意义。现有的大气河识别算法主要是基于多个物理量的阈值来实现的,它们能有效地识别大气河,但需要手动设置阈值和几何度量标准。由于现有方法在复杂度和泛化能力上的局限性,本文基于深度学习的方法,提出了名为SAR-UNet的网络模型,该模型基于UNet网络结构,在特征提取时创新性地加入了空间注意力机制,从而能够更好地获取大气河的局部特征信息。通过深度学习的方法来对大气河进行自动识别,有效地解决了现有方法需要确定阈值和度量标准的不足。实验结果表明,基于SAR-UNet的大气河识别方法在ERA-Interim气候再分析数据集上具有最佳的识别效果,大气河识别的精度上达到了97.27%,MIoU值高达85.22%。

Accurate and efficient identification of atmospheric rivers can effectively prevent natural disasters such as floods and droughts, and is of great significance to the economic, social and ecological development of the region. Existing atmospheric river identification algorithms are mainly based on the thresholds of multiple physical quantities, which can effectively identify atmospheric rivers, but need to manually set thresholds and geometric metrics. Due to the limitations of existing methods in complexity and generalization ability, this paper proposes a network model named SAR-UNet based on deep learning, which is based on UNet network structure and innovatively adds spatial attention mechanism during feature extraction, so as to better obtain local feature information of atmospheric rivers. The automatic identification of atmospheric rivers by deep learning effectively solves the shortcomings of existing methods that need to determine thresholds and metrics. The experimental results show that the atmospheric river identification method based on SAR-UNet has the best identification effect on the ERA-Interim climate reanalysis dataset, the accuracy of atmos-pheric river identification reaches 97.27%, and the MIoU value is as high as 85.22%.