Comparison of Object- and Pixel-Based High Resolution Image Classification under Different Reference Sampling Schemes

In recent years, with the development of remote sensing techniques, image classification based on remote sensing imagery with high spatial resolution has played a significant role. At present, there are mainly two ways to classify the imagery of high resolution: pixel-based method and object-based method. Because of the wide differences of spectral features in pixel-based method, 'salt-and-pepper' effect appears in the classification result, which decreases the accuracy. And the object-based method which is widely used in image classification has a low accuracy of classification about objects with similar spectral features and shape features. Each of the two methods has its advantages and disadvantages. It is generally agreed that object-based classification has an advantage over pixel-based classification. However, the previous study found that the results of object-based classification was deeply influenced by reference sampling method So it is necessary to have a comparison between object and pixel-based method under different reference sampling schemes. In this study, we studied pixel-based method and object-based method, selected samples using 'select samples randomly' method and 'select samples separated by objects' method respectively, and built classification models using SVM and RF classifiers respectively. We compared different classification methods and analyzed the impacts of selection of samples and classifiers on classification results. The research showed that the accuracy of classification depends on the distribution of sample points. When sample points were selected randomly(Rand), object-based method got a higher accuracy; when sample points were selected by reference objects separately(Sep), pixel-based method got a higher accuracy.