Evaluation of a Hierarchical Classification Method and Statistical Comparison with Pixel-Based and Object-Oriented Approaches

Aims: Producing a land use/land cover map is a fundamental step in different studies. This study aimed to assess the ability of hierarchical, pixel-based and object oriented classification methods to produce land use/cover maps. Materials and Methods: This study was conducted in the HaratMarvast basin of Yazd Province, Iran using Landsat imagery of 2016 (paths 161 and 162, row 39). The hierarchical image classification method was tested for land use/cover mapping. A statistical comparison between three algorithms, namely pixelbased, object oriented and hierarchical image classification was performed using the McNemar test. An intensive field survey was also accomplished to obtain training and test samples. Findings: The kappa coefficients for pixelbased, hierarchical and object oriented techniques were 0.76, 0.83 and 0.94, respectively. Results also showed that the performance of SVM and hierarchical algorithms are significantly different with aχ2f 112.3 which shows the superior performance of the hierarchical algorithm. Conclusion: It was shown that the object oriented approach performed significantly better than the two above mentioned methods ( χ2= 149.6). As the computational costs of object oriented methods are relatively high, the hierarchical algorithm can be suggested when there are limitations in time or computational infrastructures. Therefore, the hierarchical algorithm can be used instead of simple pixel-based algorithms for land use/cover mapping.