RECONSTRUCTING RANKINE VORTICES FROM DOPPLER WIND DATA USING DEEP-LEARNING-BASED GENERATIVE MODELS

Abstract

 To improve the accuracy of typhoon forecasts, advancing data assimilation techniques is crucial to estimate the initial conditions of typhoon forecasts. In this study, we propose reconstructing Rankine vortices (RVs), which are used as an approximation model of typhoons, from Doppler wind data. We trained two deep learning methods, autoencoder (AE) and conditional variational autoencoder (CVAE), and succeeded in reconstructing RVs from noiseless and noisy Doppler wind data. While AE provides better estimates for RVs for noiseless data, CVAE outperforms AE for noisy Doppler wind data. Sensitivity experiments to latent-space variables reveal that VAE works as a generative model for RVs.