Estimation Of Land Cover Classification Using Harmonics Analysis And Linear Spectral Mixture Model

Land use/cover classification is the basis of data and key links in land use/cover change.With the development of remote sensing science theory and deepening of the application field,The use of remote sensing technology to obtain feature target feature information has becomean important means of modern studies of its scientific research. Traditional classificationmethod based on mathematical statistics computer classification feature band, which is basedon hard statistical pixel spectral characteristics classification. For coarse spatial resolutionremote sensing data, in the vast majority of the regional scale pixel contains a number ofcategories of mixed pixels. If do this, we can’t avoid mixed pixel like synonyms spectrum orthe same spectrum of foreign body as misclassification or leakage points, The result reducethe accuracy of the estimated size of the surface feature.Vegetation index of time series is composed of discrete-time signal by the time seriesdata, showing vegetation biological characteristics of seasonal change. Vegetation index timeseries data generally requires specific analytical methods to extract feature information.Harmonic analysis of time series decompose a certain length of time series data into finiteharmonic superposition by a discrete Fourier transform. Harmonic mean, amplitude and phasecharacteristics parameters are an effective means of vegetation phenology analysis.In this paper, Hebei Plain as the study area, total of23time phases MODIS EnhancedVegetation Index (EVI)16-day250m image from January1,2011to December31,2011asthe main source of information to establish land cover classification system. Datapre-processing include data splicing, band extraction, projector and format conversion and thestudy area cutting. We use harmonic analysis of time series refactoring cloud-free images andextracting the amplitude and phase of each harmonic component. By the Minimum NoiseFraction Rotation, pixel Purity Index, n-D Visualizer determining the characteristics of themain band and using Linear Spectral Mixture Model to get the endmember abundance ratiosof one crop a year cultivated land, two crops a year cultivated land, garden and woodland,natural and artificial land surface. Take maximum value method and the dichotomy toestimate the land cover in MODIS pixel and sub-pixel scale and give a accuracy assessmentin endmember abundance ratios and totals areas. The root mean square error showed that maximum value is0.017855, minimum value is0.000091, the average value is0.000637and more than95%of evaluation results is less than0.003569. Comparing with the2010land use Statistical data, the total accuracy of garden andwoodland, water(including salt pan), artificial land surface, natural land surface in the pixelscale are71.96%,85.84%,83.57%,77.82%. Comparing the different spatial scales data oftwo crops a year cultivated land in the jurisdiction of counties (cities) of Shijiazhuang citywith field survey data, pixel scale accuracy is76.91%and sub-pixel scale accuracy is86.19%.