Comparison of machine learning and parametric methods for the discrimination of urban land cover types

Abstract

The aim of this study is to compare the performances of different machine learning and parametric techniques for differentiating highly mixed urban land cover classes in Ulaanbaatar, the capital city of Mongolia, using multisource data sets. For data sources, 17 features are chosen, including the original 10 spectral bands of the Sentinel-2 data; VV, VH, average of HH & HV, and simple ratio of Sentinel-1 data; and normalized difference vegetation index (NDVI), Bare soil index (BSI), and modified soil adjusted vegetation index (MSAVI). Six different feature combinations are used to identify available urban land cover classes. To discriminate existing classes, a support vector machine (SVM), artificial neural network (ANN), random forest (RF), and a statistical maximum likelihood classifier (MLC) are employed and compared. In all six feature combinations, the RF method outperforms the others with an overall accuracy ranging from 81.72% to 90.71%. The SVM has an overall accuracy ranging from 77.77%-83.50%, with the second-highest performance in four combinations and the lowest in two. The ANN has an overall accuracy ranging from 74.83%-83.31%, with poorer performance than the SVM. The MLC's performance varies across feature combinations, with an overall accuracy ranging from 70.96%-85.92%, and the second-highest performance in two feature combinations. Overall, the study shows that multisource information along with additional features/indices can significantly improve the classification of mixed urban land cover types, and for the given test site, the RF technique is the best option for producing a dependable land cover map.