Comparison of different machine-learning algorithms for land use land cover mapping in a heterogenous landscape over the Eastern Nile river basin, Ethiopia

Abstract

Land use/land cover (LULC) information is regarded as one of the most important variables in global change studies. Several LULC classification algorithms are available with different levels of accuracy in LULC mapping under different geographical setups. Furthermore, different remote sensing (RS) based indices and auxiliary data have been reported to have contributed to the accuracy of LULC mapping. Hence, the main aim of this study was to compare the different ML algorithms for the LULC mapping and explicitly assess the potential contribution of different RS-derived indices and auxiliary data on classification accuracy in the Upper Tekeze-Atbarah (UTA) river basin in Ethiopia. The overall classifiers evaluation result revealed that random forest (RF) performed best and followed by K-Nearest Neighbour (KNN), Support Vector Machine (SVM), Artificial Neural Network (ANN), Maximum Likelihood Classification (MLC), and Classification and Regression Tree (CART). Compared to RF, the MLC method showed less accuracy by 16.2% (19.7%) in OA (kappa coefficient) respectively, which showed that the currently emerging ML classifiers contributed to an improvement in LULC mapping. The use of RS-derived indices and DEM as predictors together with the Landsat bands improved the overall accuracy and kappa coefficient by 5.5% and 7% respectively. Furthermore, the discrepancy in the classified LULC map from RF and other classifiers was compared at pixel-level and the result showed a considerable disagreement between the LULC maps. For instance, when the LULC map from RF and KNN are compared, only 80.6% of the area has the same LULC condition. This result reflects the potential discrepancy that could exist in LULC maps from different classifiers. The hydrological simulation under each of the classified map also resulted in a considerable difference in the major water balance components. In a conclusion, given the considerable difference in classifier performance and the potential propagation of errors in LULC maps to further applications such as water resources assessment, it is highly recommended to explicitly evaluate the available classifiers. Users should also harness the power of RS indices and auxiliary datasets for improved LULC mapping.