Recurrent Neural Networks in short-term weather forecast (nowcasting) using radar images in Chapecó-SC

Abstract

In this work, a new computational approach is proposed that leverages Recurrent Convolutional Neural Networks, in which weather radar images are utilized for the prediction of the propagation and intensity of storms with up to 3 hours of anticipation, known as nowcasting. To achieve this, images from the weather radar located in the city of Chapecó-SC were used. These data are public and available on the website of the Brazilian National Institute for Space Research (INPE). To accomplish this, the utilization of a recurrent convolutional neural network with spatiotemporal learning, named PredRNN++, is suggested. The results were validated through case studies of storms that occurred within the coverage region of the utilized radar. To assess the performance of the neural network, in addition to a visual analysis of the results, Root Mean Square Error (RMSE) and Structural Similarity Index (SSIM) metrics were employed. The outcomes demonstrate that PredRNN++ was capable of simulating the shape and location where the meteorological system occurred.