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