Land cover classification of CBERS-02 images based on object-oriented strategy - A case study in Yixing, Jiangsu Province

With the rapid development of the techniques of space information, RS and computer network, the earth observation technology has also been improved. The recognition and translation of land cover with remote sensing images, as an important branch of geosciences field, has by far become a hot topic. Image information with different scales displays distinct spatial structure, so image analysis with single scale could not meet the heterogeneity and dynamic of pattern and process. Object oriented image analysis could create meaningful objects and build a hierarchical level close to surface character using multi-scale segmentation, different geographical processes could be represented in corresponding image object levels. Object oriented image analysis has realized multi-scale analysis of spatial pattern and process. The extraction of land cover classification based on objects-oriented strategy sets most priority on the multi-scale segmentation of images, the measurement of spectral, geometric and topological characteristics, the interaction between human and computer, and the construction of knowledge base. Taken China-Brazil Earth Resources Satellite (CBERS) CCD image in Aug., 2006, which is geometrically corrected via methods of quadratic polynomial and bilinear interpolation, to control the RMS within one pixel, choosing the typical urban building area and land cover abundant of Yixing City of Jiangsu Province area as the study area, and regarding Ecognition professional software as the platform, the paper carries out the classification experiment on the study areas. The paper does the research by the following steps: 1) according to the characteristics of different surface features types, choosing the optimum scale of the segmented image to extract the objects. Scale parameters are grouped into four levels, i.e. 60, 30, 20, and 5, and all the image layer weights are set to be 1, which can help to get the size of land coverage comprehensively in the area; 2) constructing the land classification system; 3) extracting the characteristics or characteristic associations of the surface feature types. The class-rules quotation of the class hierarchy and semantic structure are used comprehensively to extract the level 1 land coverage type, including cultivated land, grass land, wood land, construction land, water body, and unused land, on the basis of which, further more, the level 2 type of land coverage, including paddy fields, forest, shrubs land, urban land, rural settlements, traffic land, lake, and pond, is re-extracted via the feature-description that is characterized by the combination of multi-function and deformation function expression construction; 4) adopting integration of supervised classification and unsupervised classification to classify the surface feature types, then getting the classification result of the study area. At the end, the paper compares and appraises the classification result between the object oriented and the pixel oriented (such as the maximum likelihood classification, the ISODATA cluster classification) methods. The result indicates that :1) CBERS CCD image is a kind of moderate resolution satellite data, demonstrating object features mostly as spectral information, but little as structural information. That's why it can not exert its advantages when extracting objects during the process of image segmentation via the objects-oriented technology, as a result, there is no obvious improvement of overall accuracy of the primary class. Precisely, the overall accuracy of the objects-oriented nearest classifier is 69.89%, which is lower than the accuracy, about 71.00%, made by applying an integration of supervised classification and unsupervised classification. 2) By integrating the human-computer interactive systems, exploring the spectral information of the remote-sensing data, the semantic structure of the image, and doing more research on semantic context relationship, we get more secondary classes than by using the integrated result of supervised classification and unsupervised classification. In this way, we successfully distinguish frosts from bushes, separate traffic land from construction land, and set rivers, lakes, and ponds into different groups, by which means, the accuracy of secondary classification reaches 69.68%. 3) Object oriented classification can integrate expert knowledge and auxiliary data and overcome the salt and pepper noise.