Improving cyclone wind fields using deep convolutional neural networks and their application in extreme events

Precise forecasting of tropical cyclone-induced storm surges is necessary to avoid any significant damages to coastal communities. The numerical models need wind and pressure fields as a surface forcing to predict these events. An increase in the quality of the available wind field data greatly benefits the forecast. There remain inherent limitations in the quality of real-time wind forecasts for near-field and far-field regions surrounding the storm eye. Wind fields from global models usually underestimate the inner core winds. In contrast, parametric winds well represent the cyclone winds in the inner region. Prior studies used techniques to superpose inner and outer core winds from parametric and global models, respectively. However, for real-time deployment, such methods are computationally expensive and mathematically complex. Machine Learning (ML) techniques can be employed for real-time use once the mapping is trained to machines. In the present paper, a blending strategy that generates enhanced wind fields is trained using deep convolutional neural network architecture; it can be easily deployed once the architecture learns the mapping. Numerical simulations using ML, conventional blended winds are performed and validated against available observations. The study reveals that simulations based on ML blended winds performed as expected and are suitable for real-time operational forecasts.

Automated summary

The study demonstrates that machine learning techniques can be employed to improve the accuracy of wind field predictions for tropical cyclone-induced storm surges. The blending strategy using deep convolutional neural network architecture shows promising performance in real-time operational forecasts.