Performance Assessment of Maximum Likelihood, Random Forest and Support Vector Machines Classifier for Urban Land Use Classification: A Case Study of Dhaka Metropolitan City, Bangladesh

Segmentation of remotely sensed satellite images is obligatory for multifarious earth observation studies, including land use and land cover (LULC) analysis. It is also inherent in environment, ecosystem, and urban development in analytical perspectives and complex inputs for modeling urban planning and disaster management. Assessment of LULC pattern uses different segmentation methods for assigning specific given classes to pixels of bands containing an image of natural color composite to define land use land cover classes such as water body, vegetation, bare soil, and built-up areas. The process of assigning classes to pixels varies from one to another, and thus, different accuracy levels are obtained. The accuracy of frequently used methods for LULC classification was assessed in this study, where the Dhaka metropolitan area has been taken as a sample to observe the LULC. The classification was conducted by using three methods where the Support vector machines classification (SVMC) produced the best accuracy results of 83.2% overall accuracy and overall kappa coefficient value of 0.74 than both random forest classification (RFC) and maximum likelihood classification (MLC) methods with 86.34% and 83% spatial similarity rate respectively. Besides, RFC and MLC are roughly equivalent in kappa and overall accuracy values, though MLC revealed less capability at classifying vegetation. However, MLC showed a high spatial similarity with RFC and dissimilarity with SVMC. This study on segmentation methods in classifying LULC will help users make an informed choice in selecting the best method for relevant studies.