Abstract: Continuous and exquisite spectral signatures of surface materials are able to be obtained because of the emergence of hyperspectral remote sensing technique, which has wide spectral response range and high spectral resolution. So far, domestic and overseas researches about particle size inversion are mainly about soil, desert, etc., which have smaller range of sizes(<10 mm). Studies that focused on large gravels are few. Characteristics of Gobi surface gravel size composition reflect the process information of gobi formation, which is the premise of Gobi research and is also important in determining the complexity of gobi transformation. In order to conduct the inversion of gravel size, sensitive bands and inversion equations of gravel size were selected in combination with different forms of reflectance transformation of hyperspectral data in this study. The results show as follows:(1)the differential transform of reflectance and grain size were well correlated, and the best correlation bands were 908 nm, 983 nm and 985 nm. Among them, the reflectance after the logarithmic inverse differential transformation was positively related with particle size( R² = 0.61). However, reflectance after the first order differential, square root differential, logarithmic differential transformations were negatively correlated with gravel size, the correlation coefficients were-0.633, -0.646 and-0.649, respectively; (2)the regression analysis between reflectance data after differential transformation and particle size shows that the cubic regression model had good fitting precision, in which the logarithmic differential performed best in the regression analysis( R² = 0.851). The prediction accuracy of logarithmic differential(75.27%)was higher than the other four differential forms after validation, indicating that gravel spectra of logarithmic differential transformation of the 908 nm band could be applied to the inversion of Gobi surface gravel size; (3)the application of the optimal regressive model to predict outcomes of surface gravel size shows that the average relative error of the logarithmic differential model was 24.9% and the precision was 75.27%, while the average relative errors of the square root of differential model and the logarithmic inverse differential model were 31.3% and 33.62%, the precision were 61.82% and 59.79%, respectively, indicating that the logarithmic differential model had better application potential; (4)the gravel size distribution map obtained by applying the inversion model to the Hyperion image showed same distribution rule with field investigation data. The gravel size became smaller with the decrease of altitude. Conclusions are: the intensity of spectral absorption differed in different size gravels; hyperspectral data can be applied to obtain the size information of surface gravels because of its high spectral resolution; bands that were sensitive to the gravel size could be applied to the identification and the inversion of surface gravels in Gobi areas in combination with hyperspectral images; hyperspectral inversion models could be used to estimate the surface gravel sizes more efficiently, and it provided basis and reference for better understanding the distribution of Gobi gravels and controlling the sources of sandstorms.

Key words: hyperspectral, gobi, grain size, inversion