Comparison of object-based and pixel-based methods for urban land-use classification from WorldView-2 imagery

Nowadays, urban land use mapping has played a significant role in commerce, urban planning, tourism, as well as in environmental management. With the development of remote sensing technology, high spatial resolution satellite imagery has the capability to enable accurate classification of general urban land cover and land use types. In this study, we compared pixel-based and object-based methods for detailed land use classification, specifically, classification of commercial buildings, residential buildings and roads based on WorldView-2 imagery using support vector machine (SVM) approach. In this study we tested two classification scenarios: (1) object-based classification based on spectral, texture and geometric features in pan-sharpened multispectral image; (2) pixel-based classification based on texture features only in panchromatic image. Each scenario was trained and tested using two sets of stratified random sample (SRS) points: (1) training samples and test samples can be located within the same reference objects (SRS_1) and (2) test samples in separate reference objects from training points (SRS_2). Our results show that object-based method obtained higher overall accuracy than pixel-based method with SRS_1, but lower accuracy with SRS_2. Considering the exaggeration of accuracy with SRS_1 sample points, we concluded that for detailed land-use classification in our study, pixel-based method is advantageous over object-based method because of its higher accuracy (SRS_2) and that only panchromatic image is needed in this method.