Impervious Surface Estimation Based On Fusion Of Multi-source Remote Sensing Images

City is the most densely populated region of the human economy and life.The change of impervious surface is the most direct result of the urbanization process.A large number of studies have shown that there is a direct relationship between urban impervious surface and environmental issues.The distribution of urban impervious surface is closely related to the non-point source pollution of the atmosphere and water,and it also determines whether the city is vulnerable to flooding.Therefore,the accurate estimation of the impervious surface distribution is of great significance to the sustainable development of the environment,society and economy.Researching and analyzing the distribution and change of impervious surfaces in major cities at home and abroad are very important to the national plans and policies for urban development,environmental protection,and disaster prevention.Urban impervious region are areas where surface water permeability is significantly reduced due to human activities(eg roads,concrete pavements,building roofs,etc.).However due to the complexity and variety of land use,high-precision impervious surface estimation(ISE)is still a very challenging issue.In this thesis,a number of methods have been used to estimate the impervious surface of the city.Based on optical remote sensing,synthetic aperture radar(SAR),fusion optics and SAR methods,accurate estimates of urban impervious surface have been achieved.ISE based on fusion optics and SAR remote sensing can effectively take advantage of the complementary information of both types of remote sensing images to further improve the accuracy of impervious layer extraction..The content of this thesis mainly includes the following several parts.Firstly,ISE based on the optical image,a ground classification strategy based on gray probability histogram is proposed in this thesis,which classifies complex land use into six categories,namely,the light impervious surface,dark impervious surface,shaded area,bare soil,vegetation,water.Then,the various land types were combined into a pervious surface and impervious surface.A variety of optical feature extraction methods are applied in this thesis.In order to evaluate the distinguishing ability of these features for different objects,a feature evaluation method based on KL divergence is proposed in this thesis.Using this method can quantitatively describe the distinguishing ability of a feature for different objects.In the classification stage,this thesis proposes a method of applying deep network to the impervious surface extraction,including Deep Belief Nets(DBN)and Stacked AutoEncoder(SAE).Compared with traditional methods,the overall accuracy and Kappa coefficient by using deep network were 98.2% and 0.9647(94.81% and 0.9377 for 6 classes).In addition,the new method has a good generalization.Second,ISE based on SAR images is different from the classification based on optical image.SAR images reflect the ground scattering mechanism.Combined with this feature,A method of classification of land covers based on the H-alpha plane is proposed in this thesis.The horizontal axis represents entropy,and the vertical axis represents different types of scattering mechanisms.The ground objects are divided into five categories,namely,urban area 1,urban area 2,water,vegetation,bare soil.Similarly,the feature evaluation method based on KL divergence is also applicable to the polarization decomposition feature of SAR.By selecting suitable features,the overall accuracy and Kappa coefficient by using deep network is 98.6% and 0.9705(97.8% and 0.9724 for 6 classes).Thirdly,ISE based on fusion optics and SAR images,the key issue is the accurate registration of multi-source remote sensing images,which is different from the traditional control point registration methods.This thesis uses digital elevation model(DEM)registration method to solve the problem,where the registration error is controlled within one pixel(RMSE < 1 Pixel).In this thesis,a digital elevation model(DEM)-based SAR image registration method is used to solve the problem of registration between optical and SAR images.The registration error is controlled within one pixel(RMSE<1 Pixel).ISE based on multi-source remote sensing images is realized at the level of feature level fusion,and the overall accuracy and Kappa coefficient were 99.75% and 0.9949(99.39% and 0.9926 for 6 classes).The achievement of this thesis will help urban planners and related environmental protection agencies in the planning and construction of the city and environmental governance.In addition,the proposed method and framework can provide reference for remote sensing research in other cities.