Research On Methods For Refinement Of Impervious Cover Information

Accurate and timely information about the spatial extent and distribution of urban land cover is of great significance in monitoring the urbanization process,studying the impact of urbanization on the ecological environment and analyzing the driving force of urban development.Impervious cover is a typical type of land cover in urban areas,and many researches have been devoted to mapping impervious cover using remotely sensed data.However,the existence of mixed pixels has been proved as a key problem affecting the accuracy of per-pixel classification.Therefore,in the past few decades,the extraction of impervious cover information at sub-pixel scale has been developed rapidly.In sub-pixel classification,each pixel is regarded as a combination of its cover categories,rather than a single class label.While having achieved remarkable successes,high-quality thematic mapping of sub-pixel impervious cover still remains challenging to date.Therefore,considering the importance of impervious cover in urban land cover and the complexity of information extraction,this paper takes impervious cover as the main research class,to study the refinement of continuous percentage impervious cover(%IC)information product at sub-pixel scale,and proposes an information refinement method based on regression-kriging.The main research contents and conclusions are as follows:(1)Continuous percentage impervious cover information extraction and sample reference data collectionThe %IC information of Landsat OLI images in the research area is extracted as the %IC information product to be refined.Sampling design is carried out based on simple random sampling and stratified random sampling,and combined with sample reference data collected from GF-1 high-resolution remote sensing images,to obtain different sizes and spatial distribution of training samples and test samples,for subsequent modeling and result assessment.(2)Research on refinement of percentage impervious cover information based on regression-krigingFor the extracted %IC information product with low-quality,surface analysis is performed to extract the spatial pattern features of different neighborhood sizes as explanatory variables,and the recursive feature elimination method is used to select significant explanatory variables.Taking the %IC error(i.e.the difference between the reference value of %IC and the original value of %IC)as the response variable,multivariate adaptive regression splines,random forest and support vector regression are used to construct the regression model between the response variable and the explanatory variables.Then the residuals of the regression model with spatial autocorrelation is interpolated by kriging method,the predicted value of regression is added with the estimated value of residuals to obtain the regression-kriging estimated value of %IC error.Finally,the estimated value of %IC error is added to the original value of %IC to obtain the final refined value of %IC,thereby achieving the data refinement.The experimental results show that the above three regression-kriging methods yield refined %IC maps of improved quality,indicating the feasibility and effectiveness of regression-kriging-based refinement method of %IC information.