Scholarlink

<p>Typhoons are among the most severe natural disasters, causing significant economic losses and risks to human safety. Effective trajectory prediction is critical to mitigating these impacts. Traditional numerical methods effectively predict typhoon tracks but demand extensive computational resources. Recently, large meteorological deep learning models like “Pangu” have achieved high accuracy and efficiency using extensive reanalysis data. However, these models often neglect typhoon cloud images, which provide critical visual details, including the precise location of the typhoon eye, vital for track localization. Additionally, their high computational and data requirements limit accessibility. This study addresses these issues by constructing a multi-source typhoon dataset combining typhoon cloud images with a subset of ERA5 data used in numerical models, covering 1493 typhoon trajectory series (2005–2022) at 3-h intervals. This paper also proposes the Multimodal GAN-ConvLSTM for regional typhoon track prediction (MGCTTP), which integrates Generative Adversarial Networks (GANs) for enhancing feature generation and Convolutional Long Short-Term Memory (ConvLSTM) networks for capturing spatiotemporal dependencies in sequential data. Comparisons with WRF-ARW show that MGCTTP, using a subset of data from the WRF-ARW method, achieves superior accuracy in predicting typhoon trajectories over complex terrains, such as land and coastal boundaries, particularly for Typhoons Conson and Muifa. Moreover, incorporating ERA5 reanalysis data improves prediction accuracy compared to using only satellite cloud images. Overall, MGCTTP achieves a mean absolute error of 42.15&#xa0;km, highlighting the effectiveness of multi-source data integration in enhancing typhoon track prediction while maintaining computational efficiency. This makes MGCTTP suitable for practical applications in resource-constrained scenarios.</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

MGCTTP: a lightweight multimodal GAN-ConvLSTM framework for typhoon track prediction using satellite cloud images and ERA5 reanalysis data

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Zongsheng Zheng,
  • Jia Du,
  • Yuewei Zhang,
  • Weici Ruan

摘要

Typhoons are among the most severe natural disasters, causing significant economic losses and risks to human safety. Effective trajectory prediction is critical to mitigating these impacts. Traditional numerical methods effectively predict typhoon tracks but demand extensive computational resources. Recently, large meteorological deep learning models like “Pangu” have achieved high accuracy and efficiency using extensive reanalysis data. However, these models often neglect typhoon cloud images, which provide critical visual details, including the precise location of the typhoon eye, vital for track localization. Additionally, their high computational and data requirements limit accessibility. This study addresses these issues by constructing a multi-source typhoon dataset combining typhoon cloud images with a subset of ERA5 data used in numerical models, covering 1493 typhoon trajectory series (2005–2022) at 3-h intervals. This paper also proposes the Multimodal GAN-ConvLSTM for regional typhoon track prediction (MGCTTP), which integrates Generative Adversarial Networks (GANs) for enhancing feature generation and Convolutional Long Short-Term Memory (ConvLSTM) networks for capturing spatiotemporal dependencies in sequential data. Comparisons with WRF-ARW show that MGCTTP, using a subset of data from the WRF-ARW method, achieves superior accuracy in predicting typhoon trajectories over complex terrains, such as land and coastal boundaries, particularly for Typhoons Conson and Muifa. Moreover, incorporating ERA5 reanalysis data improves prediction accuracy compared to using only satellite cloud images. Overall, MGCTTP achieves a mean absolute error of 42.15 km, highlighting the effectiveness of multi-source data integration in enhancing typhoon track prediction while maintaining computational efficiency. This makes MGCTTP suitable for practical applications in resource-constrained scenarios.