.png){kind=icon}
.png){kind=icon}
Return
Sequence-Aware Reanalysis Encoding for Typhoon Wind Speed Estimation
.png){kind=icon}
PRE
AI
DOI：10.1109/LGRS.2025.3612296.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Although reanalysis data sequences have been widely used in tropical cyclone (TC) intensity forecasting, most existing methods still rely on single-frame or multiframe satellite imagery, or use static environmental inputs, with relatively few efforts explicitly modeling the short-term temporal dynamics of environmental factors. In this research, we explore the applicability of sequence modeling for reanalysis factors in intensity estimation tasks by proposing a lightweight fusion framework that combines a Transformer encoder for 24-h factor sequences with a Convolutional Neural Network (CNN) encoder for infrared satellite imagery. To combine environmental and structural cues, we apply a lightweight gating mechanism that re-weights temporal and spatial features before regression. In addition to standard meteorological inputs, we incorporate a smoothed lifecycle indicator that captures the storm’s developmental stage over time, enabling the model to better track temporal progression. By modeling temporal dynamics rather than relying on static inputs, our approach achieves a mean absolute error (MAE) of 3.48 m/s and a root mean squared error (RMSE) of 4.48 m/s on 183 western North Pacific typhoons from 1995 to 2003, representing an 8% reduction in RMSE compared with static-factor baselines.
Keywords:
Cross-attention
reanalysis data
sequence modeling
transformer networks
typhoon wind speed estimation
Journal
I
IF:
4.4
Papers:
579
Citations:
0