Recurrent Neural Networks applied to short-term weather forecasting using radar images from the city of Chapecó, SC, Brazil

• Autor(a) principal: Rossatto, Felipe Copceski
• Data de Publicação: 2024
• Outros Autores: Härter, Fabrício Pereira, Shiguemori, Elcio Hideiti, Calvetti, Leonardo
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Revista Ciência e Natura (Online)
• Texto Completo: https://periodicos.ufsm.br/cienciaenatura/article/view/87262
• topic: Recurrent neural networks; Nowcasting; Radar; Meteorology; Redes neurais recorrentes; Nowcasting; Radar; Meteorologia

This work proposes a new computational approach that makes use of Recurrent Convolutional Neural Networks, in which weather radar images a r e used to predict the spread and intensity of storms up to 3 hours in advance, known as nowcasting. To this end, we used images from the meteorological radar located in the city of Chapecó - SC. This data is public and available on the website of the Institute for Space Research (INPE). To this end, we propose to evaluate the use of a recurrent convolutional neural network with spatiotemporal learning called PredRNN++. The results were validated through case studies of storms that occurred in the region covered by the radar used. To evaluate the performance of the neural network, in addition to a visual analysis of the results, the MSE and SSIM metrics were used. The results show that PredRNN++ was able to simulate the shape and location of the weather system.