| As an important element of urban geographical space,buildings have important application value in land use planning,digital city modeling,disaster assessment emergency response and military target reconnaissance.With the massive acquisition and application of multi-source high-resolution remote sensing images,the practical significance of using remote sensing technology to identify the building information rapidly and accurately is becoming increasingly prominent.Object-based image analysis(OBIA)method breaks through the traditional classification of the basic processing unit in pixels limitations,to comprehensive utilization of the image object spectrum,texture,geometry and spatial structure characteristics,to achieve higher level of remote sensing image extract thematic information classification.With its obvious characteristics and advantages,it becomes an indispensable important direction and research hotspot for building object extraction.However,the surface conditions in the urban environment are complex,and different types of ground objects have their appropriate spatial scale,and it is difficult to obtain accurate image information extraction results with the uniform scale level.At the same time,there is a high degree of spectral variability in the target of the earth object,and the phenomenon of "same object with different spectrums" and "different objects have same spectrums" are prominent,and the extraction of ground object information is confronted with great challenges and interference.For building,the roof structure material is qualitative different,spectral response will also be different,form different color roof of the building,makes the performance for larger spectral differences in image,this kind of building categories of heterogeneity would also be a certain degree of influence on classification results.In addition,the comprehensive utilization of image classification need a spectral,shape and texture features,such as the number of more often,which may be irrelevant or redundant information,feature selection optimization is not reasonable,easy to cause "dimension disaster",the classification accuracy has not grown much.This paper focuses on the scale effect of remote sensing data,the spectral diversity of feature and the optimization of image classification features in the process of extracting ground feature information,and combines the data mining technology and the machine learning classification algorithm with the support of object image analysis methods.And related technical methods were studied and discussed in order to improve the accuracy of remote sensing image city building information extraction.In this paper,Zhongwei City WorldView 3 remote sensing image is used as the data source.First of all,the statistical characteristics of the different types of ground features in the study area were statistically analyzed.It was found that using multi-spectral band data can effectively distinguish water bodies,vegetation,shadows and impermeable layers and others.However,it is difficult to further distinguish the impervious ground with severe spectral confusion into buildings,roads,and bare ground.Therefore,based on the impervious physical properties of buildings and combined with the multi-scale characteristics of remote sensing images,this paper proposes a strategy for extracting buildings based on the classification of impervious layers.In order to avoid the extraction errors caused by spectral diversification of objects,the idea of classifying and analyzing the same type of ground objects based on the difference in image tone is proposed to diversify the use of different feature information of image objects.Considering the influence of feature selection on the image classification accuracy,based on the research of Relief F algorithm and particle swarm optimization algorithm,the two-stage combined feature selection method of Relief F-PSO is proposed.The research shows that the multi-level segmentation classification model and its feature space optimization method can extract city building information of high-resolution remote sensing images completely and accurately.In addition,in order to further verify the clarification of the multi-level image segmentation process,the rationality of image hierarchy construction,the effectiveness of feature selection methods and the necessity of taking into account the spectral differences.This article makes a comparative study of three classification methods in a targeted manner,and draws the following conclusions:(1)The use of a single-scale layer to extract objects of different types of images can easily cause the problem of "over-segmentation" and "under-segmentation",the classification showed serious errors and missed points.The accuracy rate of the results was reduced by 12.45%.(2)Based on expert experience knowledge and characteristics of image features,human analysis selected typical features to participate in image classification,resulting in completeness and accuracy of building target extraction,the rates dropped by 9.63% and 8.26%,respectively;(3)Irrespective of the characteristics of the spectral diversity of the objects,the overall analysis was conducted as a "category",which could easily result in the misclassification of the spectral similarities between features and the results were complete.The rate has decreased by 11.74%.(4)This method can use the minimum number of image feature information to obtain the most accurate building extraction results.The completeness rate and accuracy rate are 91.77% and 80.64%,for high-resolution remote sensing images of urban areas,the extraction of material information has important application value. |
PDF Full Text Request