Rooftop Extraction From Remote Sensing Images Based On Extreme Random Tree

With the acceleration of urbanization,there are more and more urban buildings.Buildings are very important material conditions for human survival,and they also play an important role in measuring people's living standards and living conditions.Also,as an important data source in geographic database,building information collection is particularly important.The development of science and technology constantly promotes the progress of human civilization,and at the same time,advances the development of remote sensing technology.As a very important technical means of information collection,remote sensing plays an increasingly important role in human life.In recent years,the resolution of remote sensing images is getting higher and clearer,and the information of ground objects is getting richer and richer.High Spatial Resolution Remote Sensing Image(HSRRSI)appears.HSRRSI contains abundant building information,such as spectral,geometric and texture information,which makes it possible to update the building information in the geographic database in an automated way.In this paper,we mainly start with the extraction of building rooftop information,and use the object-oriented Multi-scale segmentation(MSS)method to extract the building rooftop automatically.This paper mainly studies from the following aspects:(1)In this paper,the object-oriented method is mainly used to extract rooftop,and the object-oriented MSS method is used to segment HSRRSI.firstly,MSS parameters should be selected.in this paper,e Cognition9.0 platform and fitness function are used to evaluate the optimal segmentation scale parameters of rooftop,and find out the optimal scale parameters,compactness factor and shape factor.(2)HSRRSI contains abundant spectral,geometric and texture information.Making full use of this information is helpful to improve the extraction accuracy of rooftop,but adding too many features into it will lead to "dimension disaster",which will reduce the extraction efficiency and extraction accuracy.Therefore,this paper uses Extreme Random Tree Algorithm and Pearson correlation coefficient to carry out feature and remove redundancy,find out the feature geometry with the highest contribution and low redundancy to rooftop,and extract rooftop.(3)Sample selection of rooftop,as an important link in machine learning,occupies a very important position in machine learning.The quality of samples directly affects the results of machine learning.Good samples can improve the accuracy and efficiency of machine learning.In this paper,the nearest neighbor classification feature sample selection is used to select samples.This method first selects some samples to classify HSRRSI,and then adds the wrong image objects to the samples.This method can distinguish the image objects that are difficult to identify and classify.The extraction accuracy of rooftop can be improved.The extraction of rooftop accuracy depends not only on the selection of samples and features,but also on the fitting degree and classification accuracy of samples.Therefore,this paper selects the best classifier algorithm for classification from the discussion of machine learning algorithms,and evaluates the advantages and disadvantages of machine learning algorithms from three aspects: accuracy,fitting time and prediction time.The machine learning algorithms selected in this paper mainly include logical regression algorithm,K nearest neighbor feature classifier algorithm,decision tree classifier algorithm and random forest classifier algorithm.In this paper,the QuickBird HSRRSI data in zhanggong district is taken as the data source,and Q1 and Q2 are selected for discussion.Q1 is located at the edge of the city and belongs to the transitional zone in the process of urbanization.There are more industrial plants,less civil houses,scattered roofs and more unused land in this area.Q2 area belongs to the central area of the city,with many residential areas,dense roof distribution of buildings and close overall spatial distribution.In this paper,Q1 and Q2 regions are selected for experiments,and the best samples are selected by the nearest neighbor features.Then,the best features are selected by using the extreme random tree algorithm and Pearson correlation coefficient.Then,according to the samples and the selected features,they are added to the selected algorithms,and the classifier algorithm with the highest fitting degree with the data is found out.Finally,the classifier algorithm is selected to extract the rooftop.In this paper,the confusion matrix is mainly used to evaluate the rooftop accuracy,and the selected evaluation indexes mainly include Producer accuracy,User accuracy and Overall accuracy.The RFT extraction accuracy of Q1 and Q2 is 85.2%,97.6%,89.03% and 89.23%,91.3%,77.06% in PA,OA and UA,respectively.