EGU24-18151, updated on 11 Mar 2024
https://doi.org/10.5194/egusphere-egu24-18151
EGU General Assembly 2024
© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

Evaluating multi-task learning strategies for tropical cyclones itnensity forecasting from satellite images

Clément Dauvilliers1, Anastase Charantonis1,3,4, and Claire Monteleoni1,2
Clément Dauvilliers et al.
  • 1INRIA, Paris, France (clement.dauvilliers@inria.fr)
  • 2University of Colorado Boulder
  • 3ENSIIE, Evry
  • 4LOCEAN/IPSL

Skillfully forecasting the evolution of tropical cyclones (TC) is crucial for
the human populations in areas at risk, and an essential indicator of a storm’s
potential impact is the Maximum Sustained Wind Speed, often referred to as
the cyclone’s intensity. Predicting the future intensity of ongoing storms is
traditionally done using statistical-dynamical methods such as (D)SHIPS and
LGEM, or as a byproduct of fully dynamical models such as the HWRF model.
Previous works have shown that deep learning models based on convolutional
neural networks can achieve comparable performances using infrared and/or
passive microwave satellite imagery as input. Recently, multi-task models have
highlighted that jointly learning the future intensity and other indicators such
as the TC size with shared network weights can improve the performance in the
context of intensity estimation. This ongoing work aims to evaluate which tasks
and architectures can lead to the best improvement for intensity forecasting.

How to cite: Dauvilliers, C., Charantonis, A., and Monteleoni, C.: Evaluating multi-task learning strategies for tropical cyclones itnensity forecasting from satellite images, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-18151, https://doi.org/10.5194/egusphere-egu24-18151, 2024.