Hail causes billions of dollars in damage each year to crops and vehicles across the globe. Improving the accuracy of hail forecasting is therefore highly valuable since it can increase the benefit from timely and optimal mitigation efforts. Although there have been improved physical models for hail and a number of ways to use additional information to provide corrections to a given hail forecast, we report a novel approach to hail swath correction based on convolutional neural networks.
Convolutional neural networks (CNN’s) are a class of biologically inspired artificial neural networks that consist of a series of layers, which themselves consist of a set of learnable filters, and they have been extensively applied to visual processing tasks with great success. In this work we combine a high temporal and spatial resolution weather forecast, based on the ARW core of the Weather Research and Forecasting (WRF) model, with a CNN. For training and verification data we combine NEXTRAD radar data with four years of high-resolution numerical hindcasts of weather over the southern united states.
We demonstrate the use of and performance of a pixel-wise, multi-dimensional convolutional neural network, with stochastic gradient decent, and compare its performance and computational cost to both a logistic classification model and WRF without modification. We also define and detail convolutional networks and explain their application to the spatial prediction of hail.
Our results show that CNN’s are a promising tool to aid in the spatial prediction of hail. We discuss and demonstrate advantages of CNN’s for hail location prediction, and details of the CNN design.