Study On Urban Typical Surface Features Extraction Technology Based On Hybrid Object-based Classification Method

Thanks to the recent advances of high resolution remote sensing imagery and the land use problems in the process of urbanization,the studies of traditional cities are no longer confined to traditional qualitative analysis,and more and more effective and accurate quantitative studies have been conducted.For urban analyses with high resolution remote sensing imagery,very high spatial resolution provides a significant amount of abundant information,thereby making the information extraction based on single pixels inadequate.Object-based analyses,however,lead to a loss of spectral information due to the aggregation of pixels.Therefore,the integration of subpixel-based and object-based method provides a necessary and effective means for extracting urban areas from high resolution remote sensing imagery.Using the City of Harbin as an example,this paper adopted 2015 SPOT7 panchromatic and multi-spectral images as the study site to conduct the experiment of subpixel-based and object-based remote sensing classification method.By comparing with the classification results of traditional support vector machine,the integration of subpixel-based and object-based classification method can achieve higher classification accuracy.Major conclusions are as follows:(1)Image preprocessing methods developed in this paper mainly include geometric correction,image cutting,image fusion,and image enhancement.The fusion methods of PCA,Pansharpening,Gram-Schmidt and NNDiffuse were also employed.Results were evaluated indices of Mean,Standard deviation,Information entropy,and Average gradient.By synthesizing the direct visual evaluation and objective calculation results,the NNDiffuse is found to be the best fusion method in the study area.(2)For the subpixel-based and object-based classification of the fused image,we first developed the support vector machine to perform the pixel-based classification.Image segmentation was performed with a segmentation scale of 10 through applying the E-cognition software.With the segmentation results,a linear spectral mixture model was applied to extract four types of land covers through applying the majority voting method.Final results were obtained through systematically integrated the pixel-,object-,and subpixel-based analyses.(3)The classification results were examined through qualitative and quantitative comparative analysis.Through visual interpretation,we found that the pixel-based results have the pepper-and-salt effect,and the integration of subpixel-based and object-based classification method can effectively remove these effects.For the quantitative analyses,the overall accuracy of the developed approach is 89.08%(kappa index of 0.85),significantly higher than that of the pixel-based classification(accuracy of 84.89% and kappa index of 0.79).Therefore,the subpixel-based and object-based classification method outperforms the traditional classification method.