Comparison of accuracy and reliability of random forest, support vector machine, artificial neural network and maximum likelihood method in land use/cover classification of urban setting

Abstract

Accurate land use and land cover (LULC) is crucial for sustainable urban planning and for many scientific researches. However, the demand for accurate LULC maps is increasing; it is required to compare the classification algorithms to choose the best one. Though, machine and deep learning algorithms are widely used across the world their application is limited in Bangladesh. Accurate urban LULC mapping is challenging because urban heterogeneity affects image classification models in specific feature extraction. In this research, the accuracy of machine learning algorithms (MLA) of RF (Random Forest), SVM (Support Vector Machine), deep learning algorithm (DLA) of ANN (Artificial Neural Network) and traditional Maximum Likelihood (MaxL) method was compared in LULC classification of Dhaka city. Model accuracy of MLA and DLA was tested by statistical indices of sensitivity, specificity, precision, recall F1 etc. There is a high correlation between SVM and ANN models were found. The overall accuracy of the maps was 0.93, 0.94, 0.91 and 0.95 and kappa was 0.89, 0.91, 0.86 and 0.93 for the MaxL, RF, SVM and ANN models respectively. The user accuracy and producer accuracy largely varied according to LULC classes in the applied models. The lowest accuracy of the models was found for bare land classification followed by built-up and vegetation. The high mixture of LULC classes affects the accuracy of built up and bare land classification which produces the lowest accuracy in the MaxL model. The findings indicate that the most accurate and reliable model for urban LULC classification was the ANN model.