Scholarlink

<p>Tropical cyclone (TC) is the most destructive natural disaster in the Indian region. Its accurate intensity forecasts have posed a significant challenge to the operational agencies. In this pioneering study, we proposed a novel approach of deep learning (DL) named the Spatial Attention-based Convolutional Long Short Term Memory Network (SACLN) that takes input from various predictors to predict the categories of TCs with a lead time up to day 3. The DL model was trained and forecasted using ERA5 predictors from the years 1981–2022 over the Bay of Bengal (BoB) and validated with the India Meteorological Department (IMD) TC’s intensity datasets. Key predictors three days before the event are utilized to forecast TC from its formation, intensification to landfall every 3 h (hr) up to 72 h (Day 3). The results suggest the classification accuracy of the proposed model is 50.92% (3h), 48.16% (6h), 45.09% (12h), 40.25% (24h), 36.64% (48h), and 33.16% (72h), respectively, outperforming the skills of state-of-the-art statistical models. The recall component analysis and directional prediction errors show that the most misclassification occurs in adjacent categories, especially for rare categories such as Extremely Severe Cyclonic Storm (ESCS) and Super Cyclonic Storms (SuCS). The model results are also statistically significant and better than random guessing. Additionally, our DL model well captured the most intense and land-falling TCs, i.e., Fani, Amphan, and Nivar. The findings of this study have direct implications for effective disaster management, preparedness, mitigation, and wind engineering.</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

Improving the intensity forecast of tropical cyclones over the Bay of Bengal using spatial attention-based convolutional LSTM network

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Dhananjay Trivedi,
  • Omveer Sharma,
  • Sandeep Pattnaik

摘要

Tropical cyclone (TC) is the most destructive natural disaster in the Indian region. Its accurate intensity forecasts have posed a significant challenge to the operational agencies. In this pioneering study, we proposed a novel approach of deep learning (DL) named the Spatial Attention-based Convolutional Long Short Term Memory Network (SACLN) that takes input from various predictors to predict the categories of TCs with a lead time up to day 3. The DL model was trained and forecasted using ERA5 predictors from the years 1981–2022 over the Bay of Bengal (BoB) and validated with the India Meteorological Department (IMD) TC’s intensity datasets. Key predictors three days before the event are utilized to forecast TC from its formation, intensification to landfall every 3 h (hr) up to 72 h (Day 3). The results suggest the classification accuracy of the proposed model is 50.92% (3h), 48.16% (6h), 45.09% (12h), 40.25% (24h), 36.64% (48h), and 33.16% (72h), respectively, outperforming the skills of state-of-the-art statistical models. The recall component analysis and directional prediction errors show that the most misclassification occurs in adjacent categories, especially for rare categories such as Extremely Severe Cyclonic Storm (ESCS) and Super Cyclonic Storms (SuCS). The model results are also statistically significant and better than random guessing. Additionally, our DL model well captured the most intense and land-falling TCs, i.e., Fani, Amphan, and Nivar. The findings of this study have direct implications for effective disaster management, preparedness, mitigation, and wind engineering.