Object-Oriented Classification of High Resolution Satellite Image for Better Accuracy

Compared with the middle and low resolution satellite images, the high resolution satellite images have richer spatial but less spectral information. When these images are applied in classification and land cover-use extraction, previous study has already shown that the conventional pixel-based statistical methods can not obtain very satisfying result. To solve the problem, the paper try to introduce and test the object-oriented method based on segment to classify high resolution remotely sensed data: an image is subdivided into separated regions called objects according to the spectral and spatial heterogeneity in the image segmentation process, then the objects are assigned to a specific class according to detailed description of the class in the classification process. Regarding a QuickBird image of Wuhan as an example and Erdas Imagine, eCognition softwares as the platforms, the paper carries on two kinds of supervised classification: based on pixel and object-oriented independently, using the error matrix to analyse and compare the final classification accuracy. The experiment demonstrates that when choosing the adequate samples and segmentation parameters, the object-oriented method greatly lighten the noise influence, has higher classification accuracy and efficiency than that achieve by pixel-based method. Meanwhile, the classification result of object-oriented method is much easier to understand and explain.