Inteligência artificial na classificação de uso e cobertura da terra no semiárido de Pernambuco

Abstract

The Brazilian Tropical Dry Forest, known as Caatinga, is located in Brazil's northeastern region and has severe climatic characteristics, with dry weather and poorly distributed rainfall. Those climatic characteristics make Remote sensing analysis difficult due to its large vegetation differences between the dry and rainy periods. In order to help the remote sensing analysis in this biome, this work aims to test different Artificial Intelligence algorithms through supervised classification and to identify land use and land cover patterns in the city of Petrolina, in Pernambuco. Three algorithms were tested: Random Forest, Artificial Neural Networks, and K-Nearest Neighbors using QGIS and RStudio software based on Landsat 8 images from the dry period. Twenty samples from the classes were selected: Water, Agriculture, Urban Area, Forest, and Exposed Soil, and these samples served as a basis for training the algorithms for the classification of images. Occupancy data and precision quality assessment were obtained using Mapping Accuracy and Kappa Index, respectively: 0.9878706 and 0.9653555 for Random Forest; 0.9199973 and 0.9454833 for Artificial Neural Networks, 0.9873741 and 0.9598640 for K-Nearest Neighbors, all being considered excellent. These values were higher than those found in the most commonly used algorithms, as in the Maximum Likelihood algorithm. It was observed that the use of artificial intelligence algorithms could generate better results in the classification of land use in semiarid regions.