Convolutional Long-Short-Term Memory Networks (ConvLSTM) for Weather Prediction using Radar and Satellite Images

Nícolas de Araújo Moreira; Rubem Vasconcelos Pacelli; Yuri Carvalho Barbosa Silva; Tarcísio Ferreira Maciel; Ingrid Simões; João César Moura Mota; Cerine Hamida; Rodrigo Zambrana Prado; Emilie Caillault; Modeste Kacou; Marielle Gosset

doi:10.20906/CBA2022/3299

Nícolas de Araújo Moreira Federal University of Ceara, Fortaleza
Rubem Vasconcelos Pacelli Federal University of Ceara, Fortaleza
Yuri Carvalho Barbosa Silva Federal University of Ceara, Fortaleza
Tarcísio Ferreira Maciel Federal University of Ceara, Fortaleza
Ingrid Simões Federal University of Ceara, Fortaleza
João César Moura Mota Federal University of Ceara, Fortaleza
Cerine Hamida Massachusetts Institute of Technology, Cambridge
Rodrigo Zambrana Prado The Weather Force, Toulouse
Emilie Caillault Université du Littoral Côte d’Opale, Dunkerque
Modeste Kacou Institut de Recherche pour le Développement, Toulouse
Marielle Gosset Institut de Recherche pour le Développement, Toulouse

DOI: https://doi.org/10.20906/CBA2022/3299

Keywords: Artificial intelligence, machine learning, ConvLSTM, image processing, weather forecasting

Abstract

Artificial Intelligence techniques, mainly machine and deep learning ones, are becom- ing the most common approach for data prediction. In this context, using these techniques instead of approaches based on classical statistics has shown interesting and important contri- butions to weather prediction. The present paper discusses the prediction of rainfall and clouds direction based on a sequence of 10 and 14 frames of radar images with a loss inferior to 0.06. Two different Convolutional Long-Short Term Memory Networks configurations were tested and this work presents the estimated frames resulting from these algorithms and presents comparisons between them, real data, and with the performance of other works. The results show that these algorithms can be suitable for short-term weather forecasting.