Urban Impervious Surface Extraction Based On Multispectral-hyperspectral Remote Sensing Image Fusion

Impervious surface area(ISA)is an artificial surface where water cannot penetrate into the earth’s surface.It is an important type of urban land cover.It can intuitively reflect the urbanization process of a region.ISA proportion is an important index to evaluate the quality of urban ecological environment.With the increase of urban population and the increasing demand for living and working space,the urban land area is expanding,and the proportion of ISA is gradually increasing;In contrast,farmland,grassland,forest and other natural land are rapidly decreasing.The increasing proportion of ISA will bring a series of urban ecological problems,such as: ISA will increase surface runoff,resulting in frequent urban waterlogging;The natural surface substitution effect brought by the increase of ISA will aggravate the urban heat island.Therefore,the study of high-precision extraction of urban impervious surface is of great significance for urban ecological environment,land resource management and urban sustainable development.The spatial,spectral and temporal resolution of sensors has been enhanced due to the development of remote sensing technology,which makes it possible to extract urban ISA in a large range,fine and high frequency.Traditional ISA extraction studies mostly focus on low and medium resolution multispectral images(such as Landsat series with30 m resolution).Such studies mainly focus on the changes of ISA in macro scale and time dimension,so as to monitor the process of urbanization.However,with the acceleration of urbanization and people’s increasing attention to the urban ecological environment,the fine extraction of impervious surface has higher requirements.In recent years,a series of high-resolution optical satellites have been launched in China and abroad,such as Worldview,Gaofen-1,Gaofen-2,etc.,which successfully solved the requirements of high-resolution for ISA extraction.However,such satellites can only provide meter resolution multispectral images,which are composed of 4 ～ 8 bands and can not form a continuous spectrum,which has a natural disadvantage for the highprecision extraction of ground objects.In contrast,hyperspectral image is composed of hundreds of bands,which can form the continuous spectrum of the surface.It is an ideal data for high-precision extraction of impervious surface.However,due to energy conservation,under the same imaging conditions,spectral resolution and spatial resolution cannot be achieved at the same time.In order to solve this problem,this paper launched a research on the high-precision extraction of urban ISA based on hyperspectral remote sensing images This paper first proposes an optimization method of endmember extraction based on K-SVD to improve the accuracy of endmember extraction from hyperspectral imagery.Secondly,for the low spatial resolution of hyperspectral images and the difficulty of fusing images from different sources in practical applications,a hyperspectral-multispectral image fusion method based on endmember spatial information is proposed.Then,for the low accuracy of ISA extraction from fusion image,this paper proposed a three-dimensional convolution neural network(3D-CNN)for classification and obtain the ISA distribution map.Finally,this paper proposed a CNN-based classification method for hyperspectral spatial feature enhancement to solve the problem of poor accuracy of ISA extraction caused by low signal-to-noise ratio(SNR)features of hyperspectral images.The specific research work of this paper is as follows:(1)An optimization method of endmember extraction from hyperspectral remote sensing images based on K-SVD was proposed.For the problem of hyperspectral data redundancy and low endmember extraction accuracy,an optimization method of endmember extraction based on K-SVD is proposed in this paper.The mathematical model of mixed pixel unmixing is consistent with sparse representation theory.In this paper,the dictionary learning method in sparse expression is applied to the endmember extraction of mixed pixel unmixing,and the pixels are screened based on the initial abundance,so as to eliminate some noise and reduce redundancy of data.This method improves the extraction accuracy of endmember and helps to improve the ISA extraction accuracy.Experiments show that this method can improve the accuracy of endmember by 15.1% ～ 55.7%.(2)A hyperspectral-multispectral image fusion method based on endmember spatial information was proposed.Based on the assumption that hyperspectral and multispectral images in the same area should have the same endmember.Therefore,according to the normalized difference between hyperspectral simulated near infrared band and multispectral near infrared band,this paper masked the areas with large differences in heterogeneous data,so as to obtain more accurate hyperspectral endmember.The final fusion image is reconstructed by hyperspectral endmembers and multispectral abundance.This method effectually enhances the hyperspectral imagery spatial resolution,maintains the original spectral features to the greatest extent,and provides a good data source for the extraction of ISA in the following paper.The experimental results show that this method is the best in simulated data and real data.(3)An ISA extraction method based on 3D-CNN from fusion imagery was proposed.For the low accuracy of ISA extraction from fusion image,a 3D-CNN based classification method is proposed in this paper.On the basis of fusion image,the spatial and spectral features are extracted simultaneously by 3D-CNN for ISA extraction.In order to avoid the influence of shadow,the objects covered by shadow are labeled as training samples,which is benefit for high accuracy ISA extraction.This method can effectually distinguish the different ground objects and obtain the higher accuracy of ISA.The experimental results show that the overall accuracy and kappa of 3D-CNN are 96.38% and 0.9577 respectively.(4)A convolution neural network ISA extraction method based on hyperspectral spatial feature enhancement was proposed.For the problem of poor extraction accuracy of ISA caused by low SNR of hyperspectral imagery,a spatial feature enhancement CNN classification method is proposed in this paper.In this paper,multispectral images with the same resolution are used to make up for the lack of SNR of hyperspectral images,and their features are extracted by two groups of CNNs.Then the features are fused,and the third group of networks is used as a classifier to complete the ISA fine extraction.The experimental results show that the overall accuracy of the impervious surface in Foshan research area is 95.87%,and that in Wuhan research area is 96.24%.