A Study on Segmentation-based Classification Approaches for Remotely Sensed Imagery
遥感图像分区自动分类方法研究

For the sake of the complicated factor of objects in remote sensing images,"homo spectrum"and"hetero spectrum"co exist usually in remote sensed imagery.However,most traditional supervised merhods take the same classification criteria by spectral statistic properties for various objects in the same image file.This kind of processing influence the accuracy of classification,especially for those images which have the special characteristics such as,complicated scenes,or many differences between temporal and quality of images. For this reason,the authors put forward and have realized an approach for segmentation based classification to solve this problem.The primary procedures are completed by defining the interpretation area and classification manager,and improving the supervised classification algorithm using visual C++ 6.0 language program. Finally,the authors used TM image mosaiced by two scenes,which acquired in two different time for the neighborhood areas,and then implemented the segmentation based classification to do the experiments.The results for this experiment show: (1) The precision using segmentation based classification is obviously improved in comparing with the same schema for the whole image. (2) The interpretation area can be randomly chosen and easily obtained for the sub areas before classification according to the features of images. (3) This method can help users to choose the different schemas for classification according to the properties of the each sub areas freely. (4) This method provides the storage strategy for the classification results,for all sub areas can be stored in one file,or in different files respectively,while it is not necessary to create a new layer to store the file for the results. In short,segmentation based classification for remotely sensed imagery is feasible to classify the imageries which have"homo spectrum"and"hetero spectrum"properties,and to improve the accuracy of by the classification for every sub area divided according to imagery properties.