COMPARISON OF SENSITIVITIES OF ARTIFICIAL NEURAL NETWORKS, SUPPORT VECTOR MACHINES, AND DECISION TREES ALGORITHMS IN MULTISPECTRAL IMAGE PROCESSING: A CASE STUDY WITH LANDSAT 8

Abstract

This study aims to analyze the sensitivity of classification algorithms. The study area was selected as the Urla district of Izmir, and the spring and autumn LANDSAT 8 images of the region were processed simultaneously. Artificial Neural Networks (ANN), Support Vector Machines (SVM), and Decision Trees (DT) algorithms were used for classification. The tissues of the region consisting of urban, forest, greenhouse, meadow, citrus, cultivated, and burnt areas were selected for classification. The selection of these tissues was made by conducting in-situ observations using a GPS device. The results of the study were obtained by testing the sensitivity of classifications using the complex matrix algorithm. When considering the overall accuracies according to the seasons, the study conducted with December data provided 8% more accuracy than the one conducted with April data. According to the algorithms, the highest accuracy rate was achieved by the ANN algorithm with a rate of 85%. Although the DT algorithm was advantageous in terms of time spent on the computer, it was identified as the algorithm providing the least accuracy rate in classification. SVM algorithms are advantageous in large-scale classification studies in terms of both times spent on the computer and accuracy analyses. This study provides useful data for the conservation and management of regional natural resources. Analyzing the sensitivity of classification algorithms can provide a basis for similar studies and serve as a reference for future research.