Scholarlink

<p>Accurate estimation of tropical cyclone (TC) intensity is crucial for enhancing early warning systems and minimising the socio-economic consequences of extreme weather conditions. This study presents a systematic benchmarking of state-of-the-art pretrained vision architectures from the PyTorch Image Models (TIMM) library for continuous prediction of maximum sustained TC wind speed using infrared satellite imagery. The experiments employ the TorchGeo Tropical Cyclone Dataset, a large publicly accessible collection comprising 114,634 georeferenced infrared images annotated with wind speeds ranging from 15 to 300&#xa0;knots. Five representative TIMM backbones such as EfficientNet-B1, ResNet18, ResNetV2-50, RegNetY008, and InceptionNext are fine-tuned using mean squared error optimisation and evaluated under both 80:20 and 90:10 train and test split configurations. Model performance is assessed using a comprehensive set of regression metrics, including root mean squared error (RMSE), mean absolute error (MAE), coefficient of determination (<i>R</i><sup>2</sup>), explained variance score (EVS), root mean squared logarithmic error (RMSLE), mean absolute percentage error (MAPE), and median absolute error (MedAE). Among the evaluated models, InceptionNext consistently demonstrates superior predictive capability, achieving an RMSE of 4.021&#xa0;knots, an MAE of 2.985&#xa0;knots, and an <i>R</i><sup>2</sup> value of 0.985 for 90:10 train:test split. These results indicate that contemporary pretrained vision backbones can effectively capture complex cyclone structures from satellite observations, reducing prediction uncertainty and exhibiting strong generalisation across diverse storm conditions. Overall, the findings highlight the potential of TIMM-based deep learning frameworks to enhance operational TC intensity forecasting and support climate-resilient disaster preparedness strategies.</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

Next-generation Hurricane intensity estimation using enhanced TIMM architectures

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Naga Raju Challa,
  • Vissa Vyshnavi Sai Sri,
  • Bhuma Chandra Mohan,
  • Srinivasa Rao Uppala,
  • Pulipati Vamsi,
  • Shaik Abdul Gafoor

摘要

Accurate estimation of tropical cyclone (TC) intensity is crucial for enhancing early warning systems and minimising the socio-economic consequences of extreme weather conditions. This study presents a systematic benchmarking of state-of-the-art pretrained vision architectures from the PyTorch Image Models (TIMM) library for continuous prediction of maximum sustained TC wind speed using infrared satellite imagery. The experiments employ the TorchGeo Tropical Cyclone Dataset, a large publicly accessible collection comprising 114,634 georeferenced infrared images annotated with wind speeds ranging from 15 to 300 knots. Five representative TIMM backbones such as EfficientNet-B1, ResNet18, ResNetV2-50, RegNetY008, and InceptionNext are fine-tuned using mean squared error optimisation and evaluated under both 80:20 and 90:10 train and test split configurations. Model performance is assessed using a comprehensive set of regression metrics, including root mean squared error (RMSE), mean absolute error (MAE), coefficient of determination (R2), explained variance score (EVS), root mean squared logarithmic error (RMSLE), mean absolute percentage error (MAPE), and median absolute error (MedAE). Among the evaluated models, InceptionNext consistently demonstrates superior predictive capability, achieving an RMSE of 4.021 knots, an MAE of 2.985 knots, and an R2 value of 0.985 for 90:10 train:test split. These results indicate that contemporary pretrained vision backbones can effectively capture complex cyclone structures from satellite observations, reducing prediction uncertainty and exhibiting strong generalisation across diverse storm conditions. Overall, the findings highlight the potential of TIMM-based deep learning frameworks to enhance operational TC intensity forecasting and support climate-resilient disaster preparedness strategies.