[Analysis on possibilities of multi-spectral data for quantitative retrieving soil nutrition element contents]

Models for predicting soil nutrition elements content were established by regression methods. The data source was simulated multi-spectral data from reflectance spectra measured under laboratory condition. First, the reflectance spectra were resampled to the corresponding bands of multi-spectral sensors (TM and ASTER) according to their reflectance response functions. Then, the experiential models were established between measured spectra, simulated reflectance spectra (TM and ASTER) and soil nutrition element contents by stepwise multiple linear regression (SMLR) and partial least square regression (PLSR) methods. Precision of these models was tested by validation soil samples. Compared with models established by measured spectra, precision of simulated spectra models is slightly affected by spectral resolution. Simulated spectra models give good results for nitrogen (R = 0.89), phosphor (R = 0.79), and potassium (R = 0.68). The selected band range of SMLR models for soil N, P, and K are 2 000 to 2 300 nm, 1 650 to 1 800 nm and 600 to 800 nm respectively. The coefficients of PLSR models show that near infrared (NIR) band is more sensitive to nitrogen and phosphor than visible (VIS) band, while VIS is better for potassium. Good prediction performance indicates theoretically the future possibilities of multivariate calibration for soil nutrition element concentrations by multi-spectral remotely sensed images and bands character of sensors should be considered well because different element has different response.