In the object oriented classification, object information extracted from multi-source remote sensing imagery can be integrated to improve the accuracy and reliability of acquired remote sensing information. Before extraction information from multi-source image, the objects should be geometrically located on them. The geometrical location of objects mainly consists of three steps: the extraction and representation of objects, the pre-processing of different spatial resolution images and the transferring of objects. Each of these steps may lead to uncertainties. This paper investigates and analyzes the geometrical uncertainty of object and its influence on the extracted features and the classification accuracy. Different transferring methods may cause different uncertainty and the uncertainty of boundary pixels is the main source of the geometrical uncertainty. In this paper, two transferring methods of the objects are used: transferring of the raster objects and transferring of the vector objects. To analyze its influence on the extracted features, spectral features, NDVI (normalized difference vegetation index) and energy are calculated in two feature extraction methods for comparison. One method is based on all the object pixels and the other is based on the inside pixels without the boundary pixels. To analyze its influence on the classification accuracy, the extracted features are used to classify the objects with Bayes maximum likehood classifier. The results show that the influence of the geometrical uncertainty is small and inconstant. To avoid the instability caused by the geometrical uncertainty, the method of transferring the raster object without considering the influence of the boundary pixels may be more suitable in practice.
