Scholarlink

<p>Severe convective precipitation nowcasting is a challenging task in weather forecasting, which plays a crucial role in disaster mitigation, early warning systems, and water resource management. ConvLSTM and UNet are commonly used methods in this field. To further improve their predictive accuracy, this study proposed a framework that predicts precipitation using dual-polarization radar microphysical features. The framework consists of two components. The first is an Attentive ConvLSTM, which incorporates the mechanism to capture the velocity vector information of rain-bearing cloud layers and predicts future radar echo data based on historical radar echoes. The second is a three-channel UNet, enhanced with the Squeeze-and-Excitation (SE) mechanism, which predicts precipitation from radar echo data. On the test dataset, the improved ConvLSTM reduced the Mean Absolute Error (MAE) by up to 12.42% and increased the Intersection over Union (IoU) by up to 11.39%. In comparison with other advanced methods, the improved ConvLSTM demonstrates the best performance as the forecast lead time increases. The enhanced UNet improved MAE and IoU by up to 34.18% and 16.25%, and was compared with other advanced methods. Image visualization results also show improvements.</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

A Precipitation Nowcasting Method Using Radar Echo Data Based on Attentive ConvLSTM and Three-Channel UNet

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Chunjiang Liu,
  • Hua Wang,
  • Wei Quan,
  • Junyan Si,
  • Yixin Yan,
  • Zhitao Chen,
  • Yanming Lu

摘要

Severe convective precipitation nowcasting is a challenging task in weather forecasting, which plays a crucial role in disaster mitigation, early warning systems, and water resource management. ConvLSTM and UNet are commonly used methods in this field. To further improve their predictive accuracy, this study proposed a framework that predicts precipitation using dual-polarization radar microphysical features. The framework consists of two components. The first is an Attentive ConvLSTM, which incorporates the mechanism to capture the velocity vector information of rain-bearing cloud layers and predicts future radar echo data based on historical radar echoes. The second is a three-channel UNet, enhanced with the Squeeze-and-Excitation (SE) mechanism, which predicts precipitation from radar echo data. On the test dataset, the improved ConvLSTM reduced the Mean Absolute Error (MAE) by up to 12.42% and increased the Intersection over Union (IoU) by up to 11.39%. In comparison with other advanced methods, the improved ConvLSTM demonstrates the best performance as the forecast lead time increases. The enhanced UNet improved MAE and IoU by up to 34.18% and 16.25%, and was compared with other advanced methods. Image visualization results also show improvements.