Scholarlink

<p>Deep learning radar echo extrapolation is one of the main methods for rainfall nowcasting. However, current research neglects the inadequacies of traditional nowcasting models and overlooks the importance of input data. The inference performance of traditional nowcasting networks is observed to be poor, failing to provide accurate radar echo extrapolation results. Traditional radar systems incur large errors in describing meteorological echoes, resulting in inaccurate depiction of the current real meteorological state by the input radar echo data for deep learning networks and leading to significant inference errors. To address these challenges, we propose a novel neural architecture that integrates ConvNeXt blocks with batch normalization layers, achieving significant performance improvements over state-of-the-art baselines. Specifically, compared to FURENet, our framework demonstrates a 17% relative improvement in the critical success index (CSI). Additionally, it outperforms Video Swin-Unet by achieving 94.6% faster training convergence and obtaining higher structural similarity index measure (SSIM). Strategic incorporation of polarimetric radar variables into ConvNeXt-Unet-X2 reveals nuanced performance trade-offs. Although dual-polarization inputs (Z<sub>H</sub>+Z<sub>DR</sub>+K<sub>DP</sub>) provide only marginal improvements in SSIM (+ 0.02%) and CSI (+ 0.96%) compared to single-polarization (Z<sub>H</sub>-only) configurations under routine precipitation conditions, their significance becomes evident during extreme rainfall events. Multivariate data fusion achieves an average metric improvement of 23.61% across evaluation criteria under extreme conditions, including a 50.01% CSI increase and 32.79% false alarm rate (FAR) reduction. These results highlight the importance of polarimetric radar variables in producing predictions that more closely match observed data.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhanced rainfall nowcasting using deep learning models incorporating polarimetric radar data

  • 期刊论文

  • 发表日期：

    2025年8月18日

查看全文
  • Qiming Cheng,
  • Jiayue Zhu,
  • Yihong Su,
  • Ye Rao,
  • Fei Liu,
  • Shaochun Yuan,
  • Yang He,
  • Zhen Liu,
  • Yili Zheng,
  • Gang Zhu,
  • Yao Chen

摘要

Deep learning radar echo extrapolation is one of the main methods for rainfall nowcasting. However, current research neglects the inadequacies of traditional nowcasting models and overlooks the importance of input data. The inference performance of traditional nowcasting networks is observed to be poor, failing to provide accurate radar echo extrapolation results. Traditional radar systems incur large errors in describing meteorological echoes, resulting in inaccurate depiction of the current real meteorological state by the input radar echo data for deep learning networks and leading to significant inference errors. To address these challenges, we propose a novel neural architecture that integrates ConvNeXt blocks with batch normalization layers, achieving significant performance improvements over state-of-the-art baselines. Specifically, compared to FURENet, our framework demonstrates a 17% relative improvement in the critical success index (CSI). Additionally, it outperforms Video Swin-Unet by achieving 94.6% faster training convergence and obtaining higher structural similarity index measure (SSIM). Strategic incorporation of polarimetric radar variables into ConvNeXt-Unet-X2 reveals nuanced performance trade-offs. Although dual-polarization inputs (ZH+ZDR+KDP) provide only marginal improvements in SSIM (+ 0.02%) and CSI (+ 0.96%) compared to single-polarization (ZH-only) configurations under routine precipitation conditions, their significance becomes evident during extreme rainfall events. Multivariate data fusion achieves an average metric improvement of 23.61% across evaluation criteria under extreme conditions, including a 50.01% CSI increase and 32.79% false alarm rate (FAR) reduction. These results highlight the importance of polarimetric radar variables in producing predictions that more closely match observed data.