Long-term Inversion And Monitoring Of Impervious Surface Percentage In Mountainous Cities Based On Multi-source Remote Sensing Data

Impervious surface percentage(ISP)is an important indicator reflecting the degree of urbanization and has an important impact on the ecological environment.Monitoring long-term ISP changes in mountainous cities can provide theoretical support and decision support for optimizing development planning and strengthening environmental protection.Remote sensing digital image processing is an important means to realize fast inversion and accurate monitoring of ISP.However,the mixed pixel and spectral confusion caused by complex terrain in mountainous cities will affect the extraction accuracy of impervious surface(IS),thus reducing the accuracy of ISP remote sensing inversion.The lack of reference data caused by the lack of high resolution reference images in the early years is also a major challenge for long-term ISP inversion.Therefore,this paper proposes a method of long-term ISP inversion in mountainous cities based on multi-source remote sensing data.First,based on Landsat,Sentinel-2,Sentinel-1 imagery and terrain data,the classification features are calculated and extracted,and the Relief F algorithm is used to optimize the features of multi-source data.Using the optimal feature subset and random forest classification(RFC),the IS for the year 2020 is extracted,and further estimate the ISP reference data.Then,based on the free land cover(LULC)products and ISP reference data,the areas with invariant land-cover type are identified,and the sample sets of years(1990,2000,2010)without reference images are constructed.Finally,combined with the regression feature set and the sample set of each year,the random forest regression(RFR)models are constructed to realize long-term ISP inversion.The proposed method was implemented based on GEE platform and Java Script language.The ISP data set with a resolution of 30 m from 1990 to 2020 in the central Yunnan urban agglomeration was obtained by inversion,and the ISP spatio-temporal variation characteristics were further analyzed.The main contributions of this study include:(1)The IS extraction method based on multi-source remote sensing data reduces the impact of terrain on ISP inversion results,improves the IS extraction accuracy of mountainous cities,and ensures the accuracy of the estimated ISP reference data.(2)The method of invariant pixel recognition based on free LULC products is proposed to efficiently generate invariant sample sets.It improves the efficiency and objectivity of the identification of areas with invariant land cover and solves the problem of ISP inversion for years without reference images.The results show that:(1)Using multi-source data and RFC algorithm to extract IS of mountain cities has high classification accuracy,the overall accuracy reaches 0.96,and the Kappa coefficient reaches 0.86;(2)The long-term ISP inversion method based on the RFR models can reduce the influence of terrain and improve the inversion accuracy.The root mean square error range is 18.21%–18.51%,and the pearson correlation coefficient range is 0.53–0.68;(3)The area of IS in the study area has shown a trend of rapid growth since 1990,with a total increase of 2,250.05 km~2,an expansion of nearly 2.3 times.The average annual growth rate reaches 75 km~2/year,with the highest growth rate from 2010 to 2020.(4)The distribution of IS in the study area has obvious regional differences and terrain orientation.The IS was concentrated in the dam area with a slope lower than 15°and an altitude of 1,500–2,300 m.The landscape aggregation degree of urban areas with higher ISP continued to increase,and the landscape of urban areas with lower ISP tend to be fragmented.