Acurácia das métricas de validação da classificação de imagens

Abstract

Remote Sensing is an important tool on acquisition of information related to Earth and to accomplish many studies and consequently decision making. The data must be accurate to avoid irreversible consequences. On the usage and coverage of soil, the thematic accuracy evaluates the concordance between classification and true terrestrial, usually represented by a confusion matrix, then, accuracy index are applied, such as: total accuracy, Kappa, and others. Assuming the data is accurate, does those index offer reliability on the accuracy analysis of the maps? This paper has the objective to analyse the reliability of five accuracy index: total accuracy, Kappa, Scott's Pi, Tau and Pabak. To analyse it, was created maps of reference with four, five and six classes and maps of classification with attributed accuracy of 50%, 70%, 85% and 95%. After that, the validation was made considering that the maps are real. It was done with the purpose of compare the calculated accuracy index with the accuracy index attributed on the classification map. To validated it, were utilized windows of size 5x5, 20x20 and 25x25 pixels on random and systematic sampling on software Dinamica EGO 5, the maps were sweep to analyse the behavior of the calculated accuracy. The analysis consists on dispersion measure and central tendency of the data, histograms and regression analysis. The results shown that the most reliable index not vary on the type of sampling, but can be influenced by the number of class as well as by the type of map, in addition to the lower accuracy values such 0,50 e 0,70 tend to suffer greater variations than higher accuracy regardless of the type of index.