Predição de classes sociais com modelos de aprendizado profundo a partir de imagens de satelité e dados de renda do censo

Abstract

Socioeconomic surveys that collect information on the income or economic situation of families in the Brazilian territory are costly surveys that demand time to be carried out, for example, the Brazilian Demographic Census is carried out every 10 years, at least, which can make the formulations of inefficient public policies since, in recent years, significant population growth in cities has been notorious. However, the present work has the objective of studying and analyzing the possibility of classifying the areas of the cities of Manaus (AM) and S˜ao Paulo (SP) according to social classes through deep learning models with satellite images of the cities and data from the last available Population Census. For this, satellite images of each city were collected and, based on income data from the 2010 Census, classified into social classes A, B, C, D, and E, with this, Computational Vision models were trained for classification problems with the EfficientNetV2 architecture using the transfer learning technique for training. Finally, it obtained 16 deep learning models, 8 from the city of Manaus and 8 from the city of S˜ao Paulo, however, the 2 best models from Manaus obtained the F1-Score of 0.66 for Model 1 and 0.48 for Model 2, while the 2 best models from S˜ao Paulo obtained F1-Score results of 0.53 for Model 1 and 0.58 for Model 2.