Aplicação de técnicas de aprendizado de máquina em imagens de satélites meteorológicos

Abstract

The advance of global warming is making severe weather events more and more frequent. Weather services are essential to ensure the safety of inhabitants of the area affected by weather events such as tornadoes, hurricanes and floods. The quality of weather forecasting depends heavily on the quality and quantity of weather data. A large part of the data collection happens through weather satellites. These devices provide real-time atmospheric data to meteorological centers. The recent development of machine learning algorithms to perform data analysis has surpassed several classic algorithms in several areas of activity. The data analysis currently performed in the meteorology area is heavily based on classical algorithms described by manually programmed rules. Machine learning-based data analysis methods have the potential to improve the quality and processing speed of this data. This work explores three techniques based on neural networks for the processing of data from geostationary meteorological satellites. Such techniques are compared with the methods currently used to determine the performance of the proposed methods. The techniques explored in this work are the interpolation of image sequences, colorization of grayscale images and generation of a precipitation map using only geostationary satellite data.