Procedimentos de aprendizagem de máquina para análise de padrões espaciais com o uso da plataforma KNIME

Abstract

Satellite images from remote sensors appear as viable and efficient alternatives in the study of information on patterns of land use and coverage. Data extracted from satellite images are currently used in machine learning, and this method is able to predict the class of new data in the domain in which it was trained.Thus, the present study aimed to analyze the capacity of machine learning algorithms to predict land use and land cover in the municipality of Dona Francisca - RS. The geographic database was implemented in the QGIS software, where the import of TM/Landsat 5 images began in 2004 and 2009 and OLI/Landsat 8 for 2015 and 2019. Subsequently, the synthetic composition of the false bands RGB color 543 from Landsat 5 and RGB 654 Landsat 8, in order to obtain the samples of the reference pixels, taking into consideration the spectral information of each pixel (numerical value), in order to obtain information to characterize and differentiate patterns of land use and coverage (water, agriculture, countryside, forest and exposed soil). After the training and testing of the algorithms started in the proportions of 80% - 20%, 70% -30%, 60% -40% through the machine learning algorithms Random Forest (RF), Support Vector Machine (SVM) , K-Nearest Neighbors (KNN) and Naive Bayes (NB) in the KNIME software, and finally presented the performance of the global accuracy and the Kappa index. The results showed that the RF and SVM machine learning algorithms showed the best performances for the years 2004 and 2009. As for the year 2015, the KNN and RF algorithms had a better overall accuracy. The NB algorithm showed lower performance in all tests than the other studied algorithms. The Kappa index values generated by the KNIME software indicate that the quality of the classifications generated by the RF, SVM, KNN and NB algorithms for all years were from very good to excellent. It is evident that the machine learning algorithms showed satisfactory results, so that they were efficient in predicting land use and land cover from data from orbital images.