Research On Automatic Extraction Of Residential Area Based On Superpixel

With the increasing of population in recent years,the scale of rural residential areas is changing constantly.How to monitor the change of rural residential land timely and accurately has becoming an urgent problem to be solved.Compared to the field measurements,it has an incomparable advantage that using high-resolution images to extract residential areas,which don't be limited by terrain,altitude and other natural conditions.However,the higher the image resolution,the richer the texture and color of the road,trees and buildings in rural areas.Although the details of rural residential land information can be clearer,but it leads to traditional residents extraction method based on spectral classification can not be effectively extract residential areas from high-resolution images.In order to extract the residential area from high-resolution images effectively,this thesis proposes a fully automatic extraction method based on superpixels.This method firstly divides the image into a series of super pixels with similar features by using the super pixel method ETPS of the topology,that don't destroy the image boundary information.Then the color histogram of each super pixel is calculated and the Papanicher coefficients of the superpixel color histogram Bhattacharyya Coefficient(BC)characterizes the similarity between super pixels.Finally,based on the maximum similarity of the regional merging(MSRM)criterion,the super pixels are merged into the residential area and the background area under the guidance of the names.Using the proposed method,the residents were extracted from 0.5 m and 1 m high resolution satellite images.The results show that the proposed method can extract the residential area with a certain scale and verify the feasibility and effectiveness of the proposed method.The purpose of this thesis is to explore the possibility of using the super pixel method and the maximum similarity of the region merging method for the automatic extraction of high resolution images and the existing problems.In this thesis,we firstly analyzes several typical super-pixel segmentation algorithms,and then select a super-pixel algorithm as the method of residential extraction which can be control easily.In the process of demographic extraction,the names of geographical names are used as seeds to minimize the need for manual interaction and to maximize the automation of residential extraction.At the same time,the similarity measure is improved by using the gradient histogram of super pixels,and the process of regional merging is improved from the background and target competition.Because of time,this thesis focuses on the process of automatic extraction of residents,and verifies the feasibility of automatic extraction of residential areas that is based on super pixels and regional merging methods.In the future,we can add more image information in the process of regional merger,and use the characteristics of the building to refine the initial scope of the residents to further study.