Analysing Classification And Segmentation Parameters Selection In High Resolution Remote Sensing Image Using Based On Object

Abstract:High resolution remote sensing image which have rich surface features information can provide plenty of data information for vector data updating、disaster monitoring、land cover monitoring and so on. Traditional method-pixel based which is used to extract information from remote sensing image,only considered spectral information.But there are many topological and spatial structure information in the remote sensing image.Because of these,it is difficult to classify from remote sensing image.The object-based image analysis method for the classification of high resolution remote sensing image has brought the gospel.Multi-scale segmentation is one of the method constantly used in object-based image analysis method,multi-scale segmentation uses different scales based on the different type of features,and establish multi-level segmentation framework.The method can reflect the reality feature scale better.Also it can have a better segmentation. However, parameters in multi-scale segmentation has directly effect on the result of multi-scale segmentation, it is a important working for segmentation parameters selection.the scale selection has been studied currently,but the research in shape and compactness is little.The main work of the paper is the quantitative research on the three parameters in multi-scale segementation.Firstly,choosing the scale where brightness mean standard devitation is maximum as the optimal scale.Then analyzing on shape and compactness,using the constant interval of maximum area corresponding ground patch as the optimal interval of shape and compactness.Realizing quantitative selection of shape and compactness. Finally,according to the segementation parameters obtained,realizing multi-level and multi-scale segementation of image. And eatablishing rule-sets for classification experiment.The result shows that,method for selection of segementation parameters in this paper can obtain a better segementation,also,it is helpful to the follow-up classification.,total classification accuracy is86.5%.