Comparison of Machine Learning Classifiers for Multitemporal and Multisensor Mapping of Urban Lulc Features

Abstract. This study compares four machine-learning algorithms comprising of Classification And Regression Trees (CART), Random Forest (RF), Gradient Tree Boosting (GTB) and Support Vector Machine (SVM) for the classification of urban land-use and land-cover (LULC) features. Using multitemporal and multisensor Landsat data from 1984-2020 at 5-year intervals for the Greater Gaborone Planning Area (GGPA) in Botswana, the aim of the study is to determine the performance of the classifiers in the extraction of different urban LULC features as built-up, bare-soil, water, grass, shrubs and forest. The results show that for mapping built-up areas, RF and SVM presented the best results with overall accuracy of 85%. Bare soil is best mapped using RF and CART with accuracy of up to 98%, while SVM and GTB were most suitable for mapping water bodies. The suitable classifiers for mapping the vegetation classes were RF for grass (94.5%), SVM for shrubland (81.5%) and GTB for forest (84.3%). In terms of class specific accuracy, RF achieved the highest performance with average overall accuracy (OA) of 95.9%, SVM (95.8%), GTB (95.6%) and CART (95.1%). The same performance pattern was observed from the F1-score, True Positive Rate (TPR), False Positive Rate (FPR) and Area under ROC curve (AUC) metrices for the class classification accuracies. The overall accuracy for the eight-epoch years were RF (87.8%), SVM (87.5%), GTB (86.4%) and CART (85.3%). To improve on the urban LULC mapping, the study proposes the post-classification feature fusion of the best classifier results.