An experimental comparison of multi-resolution segmentation, slic and k-means clustering for object-based classification of vhr imagery

With smaller pixel coverages from recent sensors, considering single pixels in the classification process has become ineffective and incapable for delineation of the characteristics of targeted land use land cover (LULC) types. Object-based image analysis (OBIA) has been recently applied to enhance the identification performance of the classifiers considering not only the spectral but also the contextual and textural information. In this study, three segmentation approaches having a different theoretical basis, namely multi-resolution segmentation (MRS), Simple Linear Iterative Clustering (SLIC) superpixel algorithm and K-means clustering, were utilized to produce image objects, from which thematic maps were generated using random forest classifier. The most widely used segmentation evaluation metrics were applied by considering manually digitized reference polygons to evaluate the goodness of the constructed segments. Quality analyses were performed on four specific LULC types (residential, industrial and public buildings, and coniferous trees) for analysing the effectiveness of the approaches. Overall, the MRS algorithm produced the most accurate results in terms of both segmentation quality and classification accuracy. On the other hand, the difference in classification accuracy varied by about 4% for the segmentation algorithms. Results confirmed the importance of quality metrics for evaluating the goodness of generated segments because a direct link was observed between the quality of the segments and the classification accuracy achieved. However, the NSR metric is not favoured since it solely considers the number of segments, not the congruity of reference and corresponding image objects.