Evaluating kernel functions of support vector machines for supervised classification of land use classes

Abstract

The aim of the study is to compare the accuracy of the kernel functions of the SVM method in terms of land use classification. The study was conducted in Abant Planning Unit within the north-west of Turkey. Supervised classification was performed using Sentinel-2 satellite image. Classification was made according to land use, and kernel functions of support vector machines method such as linear, polynomial, radial and sigmoid were used. According to the findings, the classification accuracies of the kernel functions were similar to each other. However, the sigmoid kernel function showed the highest classification success (kappa coefficient=0.775). When the confusion matrix was examined, the most accurately classified land classes were broadleaf forest, mixed forest, and other areas. Kernel functions were insufficient in classifying coniferous and degraded forests.