| In the past decades,globalization has promoted the rapid development of urbanization in the world,bringing a series of ecological and environmental impacts.Obtaining impervious surface area(ISA)through remote sensing data can help understand human development patterns,enable monitoring of land cover changes,and provide a scientific basis for studies on sustainable urban development and land use changes.Due to their differences in mapping methods and training samples,there is a certain degree of inconsistency among existing 30 m datasets,and there is still space to improve their mapping accuracy.In this paper,we propose a comprehensive mapping method based on the inconsistency regions of existing multiple datasets,using the GEE(Google Earth Engine)cloud computing platform and Landsat image data to further improve the global 30 m impervious surface area extraction accuracy.The core idea of the method is to develop a targeted strategy by focusing on the inconsistency issues among existing datasets.Specifically,we divided the global terrestrial surface with square grids of 100 km.Then we calculated inconsistencies between existing ISA datasets to classify the grids into two categories: consistent grids(A-Grids)and inconsistent grids(M-Grids).Different mapping methods are used according to the grid characteristics: for consistent grids(A-Grids),fully automated sampling is used for mapping,while for inconsistent grids,manually-interpreted samples are added on top of the automated sampling,and the efficiency of manual samples is maximized using the sample time migration method.We used two independent sets of validation sample data to evaluate accuracy.One is a global validation sample set collected every five years by manual visual interpretation.The other is a validation sample set based on an existing high-resolution built-up dataset.The results show that the overall accuracy using the first test samples is 97.87%,1.27%,3.61%,and 5.91% higher than GISA1.0,GAIA,and GAUD,respectively.The overall accuracy calculated from the second test samples is 91.31%,which is still better than the existing datasets.Comparing the classification effects of different grid types,we found that both A-Grids with automatic mapping and M-Grids by adding manual samples achieved the highest overall accuracy,98.53%,and 97.86%,respectively,which verified the effectiveness of targeted mapping.And we also confirmed the necessity of manually-interpreted training samples for mapping M-Grids through crossover experiments.Based on the proposed method,we generated GISA2.0,the global ISA dataset.This dataset exhibits better accuracy and more accurate detail information than other datasets.Based on the results of GISA2.0,we calculated the global expansion of the impervious surface area.The results show that the global impervious surface has experienced rapid growth from 1985 to 2018,with the most significant increase in Asia. |
