Object-oriented Vs. Pixel-based Method For Land-use Classification Using High-resolution Remote Sensing Images

With the development of remote sensing technology, the resolution of remote sensing image, including spatial resolution, spectral resolution, temporal resolution, and radiometric resolution, has been greatly improved, and the applications are more widely. Compared with the medium and low resolution image, the high-resolution image has the richer and more obvious spatial information, geometric structure, texture information and other details. Therefore, high-resolution remote sensing images have been widely used in land-use classification. The earliest remote sensing technology used in land-use classification is the visual interpretation of classification, followed by the computer classification of remote sensing digital image. But the traditional pixel-based classification techniques of remote sensing image can obtain an better efficiency in information extraction only for the medium and low resolution remote sensing image. For high-resolution remote sensing image, pixel-based classification methods can not effectively extract the wealth information in images, and the classification results appear to be prone to a "salt and pepper phenomenon", thus leading to a lot of useless broken polygons, so the classification accuracy is very low. In order to overcome the limitations of applications of pixel-based methods in high-resolution image, some scholars have proposed object-oriented image analysis method.The object-oriented classification method is based on image objects as the basic processing unit. Corresponding to the pixel, the object is the entity in image analysis. It is composed of a collection by the homogeneous and adjacent pixels. The size of the object is decided by the image segmentation scale and image spatial structure. Therefore, the object-oriented classification method is operating meaningful image objects, and the factors involved in the information extraction include not only the spectral information, much more importantly, the shape, texture and spatial information. However, the pixel-based analysis method is based on the basic unit-remote sensing image pixels to extract information. Therefore, only pixel spectral information was involved in the information extraction for classification.The data is "a map" image of Hunyuan County. The "a map" image contains WorldView-I, SPOT-5and QuickBird data. This paper selected a part of the QuickBird image as the source data of the study area. This paper has classified the images using pixel-based and object-oriented methods respectively, then compared the two classification results and evaluated the accuracy. Using ENVI, the image was classified by the pixel-based classification including supervised classification (maximum likelihood and minimum distance), and unsupervised classification (K-means and ISODATA). Using eCognition, the image was classified by the object-oriented classification including the nearest neighbor method and membership function method. Before classification with eCognition, the author has done a lot of experiments to select an optimal segmentation scale. Through the multiresolution segmentation of the image, this paper created an image object hierarchy. This paper adopted60and120scale to do the multiresolution segmentation, selected the appropriate image feature space (such as brightness, shape, texture, etc.) as the basis for information extraction, and used the nearest neighbor classification method and the membership function classification method respectively to extract the larger surface features (such as forest land, dryland and other grassland, etc.) on the120scale and extract the small and fragmented surface features (such as village, inland tidal flats and other woodlands, etc.) on the60scale. Comparing the results of pixel-based and object-oriented classification with visual inspection and accuracy assessment, the results showed that the object-oriented classification of high-resolution remote sensing images for land-use classification information extraction can effectively avoid the "salt-and-pepper-effect". The shape and attribute of the surface features that extracted by the object-oriented classification have a higher consistency with the actual surface features. The object-oriented classification is better, more accurate and much easier to understand and interpret than the pixel-based classification.