Estudo comparativo de abordagens computacionais para classificação de imagens de satélite da Amazônia

Abstract

Monitoring large areas of forest is extremely important even though it is very difficult. The Amazon rainforest is currently considered the largest tropical forest in the world, monitoring it with the purpose of mapping its land use patterns becomes easier when thought from satellite images. Based on this idea, deep learning neural networks represent the best way to classify large volumes of images with high precision. The present work provides a comparison among deep learning techniques applied to a data set composed of more than 100,000 images of the Amazon rainforest, using Fbeta and the time obtained by each technique in the classification as the evaluation metrics. The images were labeled with one of the four classes related to the climatic condition and with none or more classes among the options related to the land use patterns. The final conclusion was that the VGG16 network would be the most appropriate choice for solving the problem, since in addition to obtaining an Fbeta of 91.79% - ranking top three among the tested techniques it classified the test database – composed of 61191 images – in time relatively low: 92.2 seconds.