Comparison of machine learning, supervised, and unsupervised algorithms for a land cover classification in northern Mongolia

Abstract

The aim of this study is to compare the performance of machine learning approaches alongside supervised and unsupervised techniques to differentiate land cover classes in the northern Mongolia. The primary goal is to identify which of these methods can achieve the highest classification accuracy. For this analysis, we selected ten original spectral bands from Sentinel-2 data and utilized three different feature combinations. To differentiate among the available classes, we employed a support vector machine (SVM), a Mahalanobis distance classifier, and K-means clustering, assessing their relative effectiveness. In the three-band feature combination, K-means obtained the lowest accuracy at 70.08%, whereas SVM achieved 92.72%, ranking it as the most effective method. The Mahalanobis distance classifier closely followed with an accuracy of 90.36%. In the five-band combination, K-means improved its accuracy to 95.56%, surpassing earlier results, while the Mahalanobis distance achieved 95.07%, and SVM recorded an accuracy of 93.33%. In the analysis involving all ten bands, K-means again delivered the highest accuracy at 95.83%. The Mahalanobis distance classifier reached an accuracy of 93.93%, while SVM had an accuracy of 93.14%. In many cases, machine learning techniques can often outperform traditional methods. However, in this study, the traditional unsupervised technique surpassed both machine learning and supervised techniques in two cases. Thus, the results suggest that achieving high accuracy is not invariably attainable with machine learning or supervised image classification methods. In many instances, it depends on the selection of parameters, the data structure, and the radiometric properties of the objects of classes.