Thursday, 6th of September

14h30 (usual room R2014, 660 building) (see location )

Mo Yang


Title: Prediction of storm trajectories


Hurricanes are one of the most severe natural disasters that cost tremendous damage and death in each year. The forecast of the incoming hurricane’ future track can be crucial for protecting people who live in shoreline areas and reducing their economic loss. In this work, we propose a deep learning framework for hurricanes trajectory forecasting. We introduce a fusion architecture, which is composed of three separate stream networks, followed by a late fusion. The proposed fusion architecture of neural networks allows for coupling different sources and dimensions of data in parallel, including data of 1D tensors (hurricane’s history track, handcrafted features) and 3D tensors (reanalysis atmospheric wind fields and pressure fields). By adding the atmospheric fields in the same location at the previous time point, the network has shown be able to capture the dynamics. This fused network is trained to estimate the longitude and latitude 24h-forecast of hurricanes and depressions from a large database from both hemispheres (more than 3000 storms since 1979). It is demonstrated that the fused network, which integrates all sources of data, significantly lowers the forecast errors than the network that uses any single source of data. We compare our method with the existing storm tracks forecasting methods. The results show that our method outperforms a commonly used statistical method. Although we do not beat the official forecast that use an ensemble method to combine a large collection of other forecasts, we show qualitative details of the two methods’ forecast which indicates that our proposed deep learning framework can help to enhance the official forecast.

Contact: guillaume.charpiat at inria.fr
All TAU seminars: here

Contributors to this page: guillaume .
Page last modified on Thursday 06 of September, 2018 11:37:30 CEST by guillaume.