Research On The Key Technology Of Monitoring Urban Land By Remote Sensing At Regional Scale

Urban land is the main area of social,economic,political and various fields of human activities.It occupies a relatively small fraction of the Earth's surface,but changes quickly.In the process of high-speed urbanization,large populations gather towards urbanized area and the demand for urban construction land increases.Urban sprawl will change local land cover situation,but it also will affect regional,even global ecological system,leading to such environmental problems as water and soil erosion,surface water pollution,ground subsidence,etc.The country's urbanization rate is still on the increase.Timely and accurate information on urbanized areas is effective to prevent blind expansion of cities,and provide scientific basis for macro-decisions.With the implementation of regional urbanization strategy,urban land monitoring and the related researches gradually turn from single city to regional or more macro-scale region.At regional scale,urban land monitoring by remote sensing can not only obtain accurate situation of land use under a low-cost premise,but also supplement history data and supply comprehensive urban land change characteristics,providing scientific guidance to regional urban land area statistics,distribution and variance analysis.Researches have made great achievements in the urban land monitoring by remote sensing,there are,however,still some questions at regional scale: 1)In terms of data,coarse imagery with low cost,wide swath and high revisit rate is the preferred data for macroscopically monitoring urban extents at the national or continental scale.However,there are lots of mixed pixels in coarse images,the “spectral heterogeneity and homogeneity” of urban land make it easily confused with other land covers,resulting in biased estimated(overestimated or underestimated).Mapping urban extent from coarse images is challenging,more research is needed to explore the respective advantages of coarse data at regional scale.2)In terms of method,regional urban mapping is typical imbalanced data classification.The non-urban pixels are considerably larger than the urban pixels.In such situations,multi-class or binary-class classification method may generate biased results.Besides,at regional scale,sampling from large area is grueling and time-consuming work.Pure urban pixels are difficult to obtain and non-urban pixels contain nonhomogeneous spectral characteristics,which worsen the sampling difficult.Thus,it is necessary to improve the classification method,developing methods that focusing on the target class only,and obtaining accurate urban extent by a small number of samples.Aiming at the problems lied in researches on regional urban land monitoring by remote sensing,this paper takes the “regional urban land” as the research object,uses three types of commonly-used remote sensing images at regional scale,nighttime light images(DMSP/OLS NTL,NPP-VIIRS DNB),MODIS products(MOD09A1,MOD13A1)and global land cover products containing urban land(MCD12Q1,GlobeLand30)as input data,and studies the effectiveness of two methods that focusing on the target class “urban land” only(i.e.,the knowledge-derived urban indices model and the data-derived one-class classification method).The main research results were as follows:(1)In terms of data,this paper analyzes data structure and data quality of the three kinds of remote sensing data,and gives the data processing methods,including removing background noise and filtering extreme values in NPP-VIIRS DNB,image selection method and pixel quality control method for MOD09A1,generating the maximum NDVI in MOD13A1,and urban extent extraction in GLC30-2010.In the experimental phase,the effectively of the combination of different data are contrastively analyzed.The experimental results show that 1)the combination of nighttime light data and surface reflectance data can obtain a higher accuracy than using single data;2)and the nighttime light data contribute more to the urban extraction results than the surface reflectance data,which have higher separability between urban and other land cover types;3)among the two nighttime light data,NPP-VIIRS DNB has a better performance on detecting low light area than DMSP/OLS NTL,especially small urban blocks and the edge of cities.Thus,the urban extent extracted by NPP-VIIRS DNB is with higher accuracy.DMSP/OLS NTL have blooming effects,saturation in urban areas and intra-sensor calibration problems,resulting in misestimates of urban land.(2)In terms of the knowledge-derived urban indices model,according to the different input data,the present indices be divided into two categories: the impervious surface indices(NDBI?NDVI?IBI?BCI)derived from surface reflectance data and the human settlement indices(VANUI?BANI?LISI)derived from surface reflectance data and nighttime light data.Based on the available indices,this paper proposed a human settlement index LHSI(Large-scale Human Settlement Index),by combining NPP-VIIRS DNB and surface reflectance data.In the experimental phase,the paper makes a thorough analysis and exploration on the distinguishability and the reliability of the eight indices.The experimental results show that 1)the human settlement indices that combines the nighttime light data are more effective to differentiate urban and non-urban pixels;2)among them,the human settlement indices(LISI and LHSI)derived from NPP-VIIRS DNB has higher precision than the indices(VANUI and BANI)derived from DMSP/OLS NTL;3)and the LHSI index can provide more inner-city details than LISI.(3)In terms of the data-derived one-class classification method,a new one-class classification method,PUL(Positive and Unlabeled Learning),is introduced into regional urban land monitoring.Through input data analysis and method comparison,this paper thoroughly evaluate the effectiveness of regional urban extent mapping by using the PUL algorithm.The experimental results show that 1)compared with the commonly-used oneclass classifier-OCSVM,PUL has a higher accuracy rate.It use the ancillary information provided by unlabeled data,avoiding the commission errors of easily confused land cover types and the omission errors of such low light area as suburban areas and small urban blocks;2)compared with the human settlement index LISI using the same input data,PUL can provided clear urban boundary,which benefits the division of urban and nonurban areas.Besides,the PUL algorithm is insensitive to outliers and parameters and has good robustness at regional scale.PUL shows great potential to map regional urban extents in an effective and efficient manner.(4)Based on the above conclusions,this paper employed the PUL algorithm to map the urban extents of China in 2012 by using NPP-VIIRS DNB and MODIS NDVI as variables.Compared with the urban extent in the datasets MCD12Q1,the obtained binary urban map is more consistent with the ground truth and has higher accuracy.A two-level accuracy assessment show that 1)on the pixel level,the accuracy of the obtained map is very high(the kappa coefficient reaches 0.842);2)on the city level,the obtained map contains all of the municipalities directly under the central government,vice-provincial cities and prefecture-level cities,only four county-level cities are omitted.The identification accuracy reaches 99.38%.