| The remote sensing monitoring of land cover is important for ecological environment monitoring,natural resource management,sustainable development planning and so on.It uses the time series of remote sensed data,detects the spatio-temporal change based on the difference of radiance by using direct comparing or trend analysis algorithms.However,the fine spatial resolution of imagery time series can hardly be acquired for large area and various change detection methods suffered pseudo changes caused by the random errors and phenological difference.This restricts the effective development of large area land cover monitoring,so it has become a hot topic focused by the researchers.One of the important approaches is to construct the fine resolution NDVI time series dataset with the help of the spatio-temporal data fusion technology,based on the existing coarse NDVI time series and multispectral images,and detect the change using the difference of NDVI profiles to reduce the adverse effect.However,the existing unmixing based methods mainly construct the adjacent pixel set with a moving window,which may result in the underdetermined problem during the unmixing procedure.Therefore,it blocks the fine spatial resolution NDVI time series dataset construction.To solve this problem,this paper proposed the idea and algorithm that construct the adjacent pixel set using the spiral,developed the two-level mapping based large-area NDVI time series construction method and proposed the algorithm that detect land cover change using multiple shape parameters of the NDVI profile.The main research work and innovations are as follows.(1)Constructing the adjacent pixel set using the spiral.The unmixing-based spatio-temporal data fusion methods often use the moving window to select the neighborhood pixels to from the adjacent pixel set.The drawback of moving window was proved by a simulation and statistics among the globe,which often lead to the underdetermined problem when solving the fine resolution NDVI values.The Archimedean spiral(abbreviated as spiral)is a kind of curve formed by gradually rotating and expanding along the center outward.It can be used to traverse neighborhood pixels based on class components and find the pixels with correlated components.The spiral path in raster space and the necessary conditions searching the adjacent pixels were defined.An experiment in Shandong province using the Globeland30 data showed the spiral-based algorithm reduced the numbers of the underdetermined pixels from 200 to 1 in each ten thousands pixels.It adapted the complex land surface and no outlier apparent in its unmixed image.The consumed time was twice compared the moving window methods.(2)Constructing large-area NDVI time series dataset using the two-level mapping.The precise pixel correspondence between NDVI data and multi-spectral images is the premise of constructing large-area NDVI data sets.However,due to the obvious differences in scale between the decomposed NDVI pixels and the high spatial resolution pixels,it is difficult to form a complete spatial correspondence.To solve this problem,the spatial relationship of tiles and pixels between MODIS-NDVI and Landsat WRS was analyzed and the two-level mapping relation was developed.The large-area NDVI time series dataset construction was proposed,which used the two-level mapping relation,linear mixing model and object-based adaptive clustering method.Totally 644 NDVI images among the Shandong province in years 2010 and 2015 had been constructed and the ability to construct one NDVI image was about 15 minutes.The spatial resolution of the generated NDVI images was improved from 250-m to 30-m.The correlation coefficient between old NDVI profile and new NDVI profile was above 0.9,and their mean absolute differences were 0.16 and 0.2,standard deviations were 0.1,in 2010 and 2015 respectively.(3)The change trend description and change detection of the NDVI profiles based on multi shape parameters.Influenced by phenological differences,the spectral and NDVI curves of the same land object are often quite different,which brings difficulty to change detection.The change detection algorithm was mostly constructed based on the direction and angle of the time series curve.It is difficult to describe the whole change trend effectively,resulting in the low contrast of the calculated change intensity and the low precision of change detection.By analyzing the phenological characteristics of NDVI profile,the overall trend of the profiles was measured by phase angle,baseline accumulation,relative accumulation and zero-crossing rate in change detection.Four shape parameters,namely,phase angle accumulation,baseline accumulation,relative accumulation and relative baseline zero-crossing rate,were derived to describe the NDVI profiles affected by phenological differences.Then,a change detection method based on multi-shape parameters was proposed,which includes calculating time series and changing intensity of spectral single-shape parameters,normalized integration and automatic threshold-based change region discrimination.It can overcome the adverse effects of phenological differences on detection accuracy and reduce the dependence on filtering.Experiments showed that the proposed method could improve the contrast between the changed and unchanged land cover from 18% to 32%,make the changed objects more prominent in the changing intensity image,and could select the threshold in a wide range of values,and improve the accuracy of the change detection significantly(up to 88%).The comprehensive study have shown that the proposed method enhances the adaptability of large area land surface approximation,improves the construction ability of large area NDVI time series data sets,and significantly improves the contrast and change detection accuracy in complex areas.This study solves the problems encountered in the large areas land cover change detection,and is beneficial to provide technical support for the continuous monitoring of large area land surface. |
PDF Full Text Request