Extraction Of Gobi Information And Quantitative Inversion Of Surface Gravel Grain Size Based On TM Images

Gobi, widely distributed in northwest China, is the main desert landscape existing in aridregion.It has rich natural resources, great economic value, and great significance in nationaldefense, political and social aspects.However, the research data about Gobi area is scarcebecause of the severe natural condition and the difficulty using traditional survey method,which is time-consuming and laborious to conduct.The development of modern science andtechnology, especially the progress of remote sensing and spatial information technology,provided great technical support for the determination of Gobi area and distribution, as well asthe partition and recognition of the ground substance composition. Because of the scarceresearch data, which caused the indistinct of the distribution and area of Gobi and the necessityto conduct research in Gobi area, the extraction of precise Gobi area is priority.Among thevarious characteristics of Gobi, the characteristics of the ground composition affect the natureof the other features directly and determine the difficulty in rebuilding and utilization to a largeextent.The formation process of Gobi can be reflected through the grain size of Gobi surface,which helps to understand the characteristics of Gobi and analyzethe causes of natural disastersin Gobi area. The study of the Gobi surface grain size contributed great effort to cognize thelaws of sand grain movement and desert extensionand provided guidance to sand preventionengineering.The typical gobi area in the eastern part of Tianshan mountains in Hami, Xinjiang wastaken as the study area. The Landsat TM remote sensing images in2010and the30m DEMwere taken as the basic data source. The decision tree model for information extraction basedon expert knowledge was establishedfirst,which was on the basis of the analysis of differentspectral characteristics of various land types. The land use types of the images were classifiedto extract the Gobi area in a better way.There were correlation between the Gobi surface grainsize and the multi-spectral remote sensing data, the vegetation index and geological factors. However,multiple correlation problems could exist among these factors, which could bringabnormal model when the model was eatablished directly with these factors to estimate theGobi surface grain size.The Principal Component Analysis (PCA) could be used for screeningfactors to reduce the dimension and simplify the model, which could reserve the maininformation of several related factors and avoid the collinearity problem among the factors.Theprincipal components were screened from the43remote sensing/geological factors (includingband information of the image, DEM, NDVI, GEMI, the SBI, GVI and WVI obtainedthroughK-Tchange, the texture factors which were mean, variance, entropy, correlation andcontrast obtained through texture analysisand roughness extracted from DEM, etc.) using PCAmodule of ENVI software.The regression model of the Gobi surface grain size was establishedtaking the principal components as the independent variable and the grain size obtained fromfield investigation as the dependent variable in SPSS.The model had significant correlationafter the verification using variance analysis and correlation test.The Gobi surface grain sizewas estimated using the estimation model and the estimation accuracy was validated. The studycould help us to divide Gobi area precisely, have better understand of the characteristics ofGobi and contribute to the rebuilding and utilization of Gobi, regional disaster reduction andthe western economic construction.The main results were as follows:(1) the Gobi area was extracted precisely using the decision tree classification, and the unusedland and other land types were seperated, the overall classification accuracy was over90%, theKappa coefficient was0.919, the extraction accuracy of Gobi areas wasover95%.(2) the remote sensing/geological factors correlated to the Gobi surface grain size werescreened and the PCs were extracted using PCA. The contribution rate of the first five PCs was98%, which reflected the main information of the sample. The estimation model wasestablished taking the first five PCs as the independent variable and the corresponding grainsize as the dependent variable. According to the validation, the model had significantcorrelation (correlation coefficient R=0.825, significant level α=0.01).The quantitativeremote sensing inversion of the Gobi surface grain size was conducted using the estimation model.And the validation showed that the estimation value and the measured value wereclosely related (the correlation coefficient R=0.778). The estimation model was effective andprovided technical support for the research of Gobi.