Avaliação de classificadores basados em aprendizado de máquina para a classificação do uso e cobertura da terra no bioma caatinga

Abstract

Proper management of natural resources in fragile environments, such as the Caatinga, requires knowledge of their properties and spatial distribution. In this context, the study aims at evaluating the performance of two algorithms based on machine learning Multi Layer Perceptron (MLP) and Support Vector Machine (SVM)) and the Maximum Likelihood method to classify land use and land cover in the Caatinga biome. For the experiment, it was used a satellite image of LANDSAT-5/TM containing the study area located in the municipality of Iguatu-CE, and classes of land cover, namely: anthropized by agriculture, other types of anthropized, water, herbaceous shrub savanna (CHA ) and dense arboreal savanna (CAD) were defined. The performance of the methods was analyzed by the coefficient of Global Accuracy (EG), Accuracy Specific (EE) and Kappa (K) coefficient calculated with data taken from the confusion matrix corresponding to ground truth. The coefficient of EG were: 86.03%, 82.14% and 81.2% and K: 0.77, 0.76 and 0.75 in the methods SVM, MLP and maximum likelihood respectively. EE values were above 70% for all classifiers tested. The results have shown that SVM and MLP methods are suited to the classification of the proposed standards, as it showed similar results to the traditional method of maximum likelihood. However, these methods are more time consuming in the stage of defining the parameters of the network and may require more computation power during stage of processing.