Comparison of Neural Network and Maximum Likelihood classifiers for land cover classification using landsat multispectral data

Land cover classification is one of the remote sensing applications, in order to identify features such as land use by utilizing typically multispectral satellite data. Numerous algorithms have been developed for classification purpose and different classifiers have their own characteristics. Different data and study area, especially the landscape complexity bring different impact on the different classifiers. Therefore, the aim of this study is to compare Neural Network and Maximum Likelihood approaches and find a suitable classifier in land cover classification using medium spatial resolution satellite images in equatorial tropical region. These two classifiers were tested using Landsat Thematic Mapper (TM) data in Penang Island, Malaysia using the same training sample data sets. Five land cover classes - forest, grassland, urban, water, and cloud - were classified. In addition, the study also been carried out in order to obtain the performances of both classifiers for the purpose of land cover mapping. Overall classification accuracy and Kappa Coefficient were calculated. The results indicated that Neural Network algorithm provided better classification accuracy than Maximum Likelihood algorithm. The overall accuracy of Neural Network approach reaches 93.5 % associated with 0.909 Kappa coefficient, which is more reliable than Maximum Likelihood, with 80.5 % overall accuracy and 0.722 Kappa coefficient.