Remote Sensing Methods For Mapping Percentage Vegetation Cover In Desertification Areas

Percentage vegetation cover(PVC)is an important parameter reflecting the changes of ecosystems,and it has been used for monitoring and evaluating desertification.Therefore,it is of great significance to estimate the PVC in desertification areas quickly and accurately.Remote sensing is the most important ways to obtain regional and global PVC.However,due to the sparse vegetation and sparse population in desertification areas,it is difficult to collect the field observations of PVC,and mapping PVC of the areas is challenging.It is necessary to further compare and analyze the applicability of PVC remote sensing estimation methods in desertification areas.The regression methods,nonparametric methods,and spectral unmixing methods were used to mapping the PVC over Duolun County of Inner Mongolia,China,using Landsat 8 images and 920 sample plots sampled by stratified system.A total of 255 remote sensing variables were derived and their correlations with the PVC values were analyzed.Then,stepwise regression and variance inflation factor were adopted for the selection of the spectral variables.The selected variables included SR435,SR23,1/B1,SR42,DVI56,which were used as independent variables to estimate the values of PVC by linear stepwise regression(LSR),support vector machine(SVM),random forest(RF),and radial basis neural network(RBFNN).In this study,three novel methods,nonlinear spectral unmixing based on squared Euclidean distance(NLSU-SED),nonlinear spectral unmixing based on weighted k-Nearest Neighbor(NLSU-WkNN),and nonlinear spectral unmixing based on optimized weighted kNN(NLSU-OWkNN),were proposed based on spectral space distance and improved k-Nearest Neighbor(kNN)algorithm.And squired Euclidean distance were used to purify the endmembers.The accuracy and cost-effectiveness of the above three methods were compared with those of the Linear Spectral Unmixing(LSU),regression methods,and nonparametric methods which is to obtain the suitable methods for PVC mapping in the study area.The main findings are as follows:(1)In the regression methods and nonparametric methods,the random forest has the best result in estimating PVC.Random forest has the highest estimation accuracy of PVC,and the PVC estimates are within a reasonable range.The estimation accuracy of PVC obtained by LSR,SVM and RBFNN is similar to that of RF,but LSR,SVM and RBFNN produce unreasonable estimates.(2)Using squired Euclidean distance to purify endmembers can improve the PVC estimation accuracy of LSU.After endmember purification,the PVC estimation accuracy of LSU is improved from 56.8%to 67.6%,and the underestimation problem is improved to some extent,which makes the PVC estimates closer to the measured values.(3)Three nonlinear spectral unmixing methods in this paper improve the underestimation problem of linear spectral unmixing and obtain higher estimation accuracy.NLSU-SED has the highest accuracy in the nonlinear spectral unmixing methods,followed by NLSU-OWkNN and NLSU-WkNN.And NLSU-SED,NLSU-OWkNN and NLSU-WkNN improve the PVC estimation accuracy from 67.6%to 77.1%,72.5%and 67.9%,respectively,compared with LSU after endmember purification.Meanwhile,the mean PVC values estimated by NLSU-SED and NLSU-OWkNN are within the confidence interval of the mean PVC values of sample plots.(4)The NLSU-SED has the highest cost-effectiveness in this study,and it is more suitable for desertification areas.Although random forest has the highest PVC estimation accuracy,it requires a large number of sample plots and is less cost-effective than NLSU-SED,which does not require sample plots.Because of the special natural environment conditions in the desertification areas,both accuracy and cost should be taken into account.Therefore,NLSU-SED is more suitable for PVC mapping in the study area.