Scholarlink

The increasing frequency of natural disasters, particularly extreme weather events, underscores the significance of accurate and rapid weather forecasting. Traditional Numerical Weather Prediction (NWP) can take up to 12 h to generate a forecast, which is often insufficient to respond to atmospheric changes. Deep learning approaches significantly reduce processing time, once the model is adequately trained, addressing the critical intervals during which extreme weather events frequently occur. This research employs two deep learning models, in combination, to address the temporal and spatial dependencies of weather forecasting. In our approach, we utilize TimeDistributed Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks that have convolutional layers capable of handling spatial dependencies and using recurrent cell can manage temporal dependencies. This study uses ground-based observations and satellite images, focusing on an area of 2,500 km2 in the Lampung Province of Indonesia. In this research, we evaluated the combined model by comparing it with several models that have been widely used. The overall performance of TDCNN-LSTM outperformed TDCNN, Convolutional LSTM, and CNN-LSTM to predict rainfall classification for up to 12 lead hours.

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

Utilizing a Time Distributed CNN-LSTM to Address Spatiotemporal Dependencies in Weather Forecasting Using Satellite Imagery

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Achmad Raflie Pahlevi,
  • Akbar Ghobakhlou,
  • Jacqueline Whalley

摘要

The increasing frequency of natural disasters, particularly extreme weather events, underscores the significance of accurate and rapid weather forecasting. Traditional Numerical Weather Prediction (NWP) can take up to 12 h to generate a forecast, which is often insufficient to respond to atmospheric changes. Deep learning approaches significantly reduce processing time, once the model is adequately trained, addressing the critical intervals during which extreme weather events frequently occur. This research employs two deep learning models, in combination, to address the temporal and spatial dependencies of weather forecasting. In our approach, we utilize TimeDistributed Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks that have convolutional layers capable of handling spatial dependencies and using recurrent cell can manage temporal dependencies. This study uses ground-based observations and satellite images, focusing on an area of 2,500 km2 in the Lampung Province of Indonesia. In this research, we evaluated the combined model by comparing it with several models that have been widely used. The overall performance of TDCNN-LSTM outperformed TDCNN, Convolutional LSTM, and CNN-LSTM to predict rainfall classification for up to 12 lead hours.