TFA-Net: A Temporal Feature Aggregation Framework for Tropical Cyclone Intensity Estimation From Satellite Images

Deep learning models have significantly advanced tropical cyclone (TC) intensity estimation from satellite images. While the potential of temporal information for improving TC intensity prediction is recognized, existing methods have not fully leveraged this aspect. To address this, we propose TFA-N...

Full description

Saved in:
Bibliographic Details
Main Authors: Zhitao Zhao, Zheng Zhang, Qiao Wang, Linli Cui, Ping Tang
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Subjects:
Online Access:https://ieeexplore.ieee.org/document/10974485/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Deep learning models have significantly advanced tropical cyclone (TC) intensity estimation from satellite images. While the potential of temporal information for improving TC intensity prediction is recognized, existing methods have not fully leveraged this aspect. To address this, we propose TFA-Net, a deep-learning temporal feature aggregation framework based on a dual-branch Transformer, for TC intensity estimation. To enrich the available data features, our TFA-Net utilizes both image and historical intensity sequences. To enhance feature exchange within and between these sequences, the framework employs global attention tokens and cross-attention modules. Furthermore, to adaptively focus on different temporal lengths, a gated feature fusion module combines models with varying input sequence lengths, allowing TFA-Net to consider both long-term and short-term TC features for improved prediction. Experimental results show that the estimation performance is improved through our model's in-depth consideration of the temporal continuity of TC. Our model achieved a root-mean-square error of 7.21 knots on the TCIR dataset, demonstrating the Transformer's potential for real-time TC intensity estimation by effectively extracting time-series information from satellite imagery.
ISSN:1939-1404
2151-1535