The Identification And Spatial Heterogeneity Analysis Of Impervious Area From Sentinel-2 Imagery

The spatial distribution and geometric shape of the impervious surface are an important indicator for urban planning and urban environment research.In recent decades,the rapid development of aerospace technology has also provided more possibilities for urban impervious surface extraction.How to extract the impervious surface from the remote sensing image has become the focus of many researchers.This paper takes convolution neural networks as an example to discuss the feasibility of deep learning for extracting impervious surfaces by comparing the accuracy of impervious surfaces identification with models of different complexity.Then this paper analyzes the spatial heterogeneity of the impervious surface in different areas.Finally,the analysis results are applied to traditional index methods,and the new index methods and the impervious surface extraction models are combined to realize the impervious surface extraction.The main content of this article can be divided into three parts:(1)Feasibility analysis of impervious surface extraction based on deep learning method.This article takes the convolution neural network as an example to introduce the theoretical basis of deep learning model implementation,and analyzes the accuracy of different complexity models for the impervious surface extraction by building convolution neural network models of different complexity.By comparing the impact of different complexity models on the impervious surface extraction,not only can it help to compare the performance differences between these models to get the most suitable model for this work,but also reveals the sensitivity of this task to changes in the complexity of deep learning models.(2)Analysis of spatial heterogeneity of impervious surfaces in different regions.There are three research areas in this paper,which are the sparse area,medium area and dense area of impervious surface.By analyzing and classifying the distribution characteristics and material properties of impervious surfaces in different regions,the spectral feature values of different types of surface coverage corresponding to different wave bands are analyzed,and the possibility of extracting impervious surfaces in different regions by different wave bands is discussed.(3)The combination of traditional index and deep learning model.This article divides the urban surface coverage into three categories: impervious surfaces,permeable surfaces and water.In this paper,the original water index and building index are improved by analyzing the spectral characteristics of the impervious surface.Then the gap between the three types of traditional index method and the new index method in different categories in each region was compared.The improved water index,improved building index and perpendicular impervious surface index are combined with the original band as the input of convolution neural network models.And the classification results of different model inputs are compared.The experiment not only proved the feasibility of the deep learning model in the impervious surface extraction task,but also showed that the use of different indexes and the original band to form a new feature input can improve the accuracy of the deep learning model.