Analysis of Land Cover Classification Results Using ANN, SVM, and RF Methods with R Programming Language (Case Research: Surabaya, Indonesia)

Abstract

Remote sensing technology has experienced rapid development, especially in the field of image classification method. Image classification is to detect and identify objects on the earth’s surface using satellite images. Currently, various satellite image classification methods have been developed, one of which is a parameter-based classification method, such as MLC (Maximum Likelihood). However, this method cannot be performed in environments with complex object features (such as in urban areas or densely built areas). This is because the classification method based on the required parameters of the dataset is not normally distributed. Therefore, a non-parametric-based classification method has been developed that is independent of the nature of the data distribution, so that no statistical parameters are needed to separate many classes in the image and are used to analyse satellite images with dense and complex land cover features. Through this research, a comparative analysis of land cover classification was carried out using a non-parametric-based classification method with several Machine Learning approaches such as Artificial Neural Network, Support Vector Machine, and Random Forest for the Surabaya City area using Landsat-8 satellite imagery data, to find and compare their classification performance. Surabay awas chosen as the research area because it has complex land cover characteristics and is quite dense. The composition of the training points used is 80:20, which 80% points as sample points, and 20% as validation points. The land cover classification results in this study were tested for accuracy quantitatively and qualitatively. Based on the result from the accuracy test, the Random Forest method showed the best results with an overall accuracy value of 93.33% and a kappa accuracy of 91.07%.