A comparison of pixel-based and object-based classification methods, a case study: Istanbul, Turkey

Since thematic classes are represented with high spectral variance, pixel-based classifications generally result with incontinuous and inhomogeneous outputs. Objectbased classifications overcome this problem by the approach similar to human seeing and interpreting activity. First, image is segmented into smaller objects, and then image objects are assigned to classes according to their spectral reflectance statistics, shape & texture properties, and neighborhood relations. Segmentation and classification can be structured as a multi-level network. The final output is formed by obtaining the classification results from the most suitable level for each thematic class. In this study, in order to investigate an appropriate method for a general landcover mapping, isodata and maximum likelihood methods of pixel-based image classifications were compared with condition based and nearest neighbor methods of object-based classifications. Tests were applied on a multispectral SPOT 5 data of Istanbul, Turkey. The results showed that the performance of condition-based method is better than the others.