[Soil taxonomy on the basis of reflectance spectral characteristics]

Soil spectral reflectance is the comprehensive representation of soil physical and chemical parameters, and its study is the physical basis for soil remote sensing and provides a new way and standard for soil properties themselves' research. Soil room spectra significantly correlate with that derived from hyperspectral images. So the room spectra are very important for soil taxonomy and investigation. To seek for the feasibility of soil taxonomy on the basis of topsoil reflectance spectral characteristics, and provide the theory foundation for quick soil taxonomy based on remote sensing methods, the spectral reflectance in the visible and near infrared region (400-2 500 nm) of 248 soil samples (black soil, chernozem, meadow soil, blown soil, alluvial soil) collected from Nongan county, Jilin province was measured with a hyperspectral device in room, and the soil spectral characteristics were determined with continuum removal method, and soil spectral indices (spectral absorption area, depth and asymmetry) were computed, which were introduced into BP network models as external input variables. The models consist of three layers (input, output and hidden layer), the training function is "TRAINLM", learning function "LEARNGDM", and transferring function "TAN SIG". The results showed that: (1) There are some differences among different soils in their spectral characteristics, but with similar parental matrix and climate, the spectral differences of soils in Nongan county are not significant. So it's difficult to analyze soil spectral characteristics based on soil reflectance. (2) The curves after continuum removal strengthened soil spectral absorption characteristics, and simplified soil spectral analysis. The soil spectral curves in Nongan county mainly have five spectral absorption vales at 494, 658, 1 415, 1 913 and 2 206 nm, and the former two vales are caused by soil organic matter, Fe and mechanical composition, the latter three are due to soil moisture; the differences of the latter three vales among different soils are not apparent, and the significant differences are in the former two vales region. (3) Soil reflectance is sensitive to organic matter, soil moisture, Fe, mechanical composition, roughness, and so on. The sensitivity of soil spectral indices derived with continuum removing method is decreased. Then the models with these indices as input variables are more stable and general. As the input variables were external, the BP network model based on the former two vales' shape characteristics was better than that based on reflectance values or all five vales, the classifying accuracy of the main three soils (chernozem, meadow soil, blown soil) was bigger than 60%, and the model could be used for soil taxonomy. However, this work still needs further study, and to improve classifying accuracy, auxiliary data, such as topography, vegetation, and land use should be introduced.