复旦大学周文团队的一项最新研究开发出从卫星图像中提取热带气旋（TC）强度的视觉转换器。该研究于2024年7月3日发表于国际一流学术期刊《大气科学进展》杂志上。

本文提出了一种基于注意力机制架构的视觉转换器(Vision Transformer, ViT)ViT-TC模型。卫星图像包括红外(IR)、水蒸气(WV)和无源微波(PMW)，以作为TC强度估计的输入。实验表明，混合输入IR、WV和PMW比其他输入通道的组合能给出更准确的估计。研究采用集成平均技术，将模型的估计精度分别提高，均方根误差(RMSE)和平均绝对误差(MAE)为9.65和6.98节，优于传统方法，与现有深度学习模型相当。

该模型对PMW高的区域赋予较高的关注权，表明PMW的大小是模型估计的重要信息。该模型还对具有高IR和WV的非云区域给予了很高的关注权重，这表明该模型检测到了TC大小和强度之间的正相关关系，并从样本边缘的非云区域导出了这一特征。

据研究人员介绍，TC强度估计是TC观测和预报中的重要内容。深度学习模型最近被应用于从卫星图像中估计TC强度并产生准确的结果。

附：英文原文

Title: Vision Transformer for Extracting Tropical Cyclone Intensity from Satellite Images

Author: YE TIAN, Wen Zhou, Paxson Cheung, Zhenchen LIU

Issue&Volume: 2024-07-03

Abstract: TC intensity estimation is an essential task in TC observation and forecasting. Deep learning models have recently been applied to estimate TC intensity from satellite images and produce accurate results. This work proposes the ViT-TC model based on the Vision Transformer (ViT) architecture built by the attention mechanism. Satellite images of TCs, including infrared (IR), water vapor (WV), and passive microwave (PMW), are used as inputs for TC intensity estimation. Experiments show that inputting a combination of IR, WV, and PMW can give a more accurate estimation than other combinations of input channels. The ensemble mean technique is applied and improves the model’s estimations to a root mean square error (RMSE) and mean absolute error (MAE) of 9.65 and 6.98 knots, which outperforms traditional methods and is comparable to existing deep learning models. The model assigns high attention weights to areas with high PMW, indicating that PMW magnitude is essential information for the model’s estimation. The model also gives high attention weights to non-cloud areas with high IR and WV, suggesting that the model detects the positive correlation between TC size and intensity and derives this feature from the non-cloud area over the edge of the sample.

DOI: 10.1007/s00376-024-3191-1

Source: http://www.iapjournals.ac.cn/aas/en/article/doi/10.1007/s00376-024-3191-1viewType=HTML