Determination Of Optimal Hyperspectral Variables To Monitor Wheat Biomass

It is critical to estimate the biomass for assessing crop growth and predicting yield in crop. The hyperspectral techniques provide a powerful technique for monitoring crop biomass. The previous studies about using hyperspectral data to study crop mainly focused on models based on the full spectra or the manually selected spectra. The stability and prediction ability of full spectra models may be weakened because of involving noises, other unrelated and collinear spectral variables. The manually selected spectra were extracted by vegetation indices, spectral absorption features, derivative spectra and spectral locations in common use, which may ignore the other spectral information, not identify the high biomass and impact the accuracy of model. In order to extract the optimal hyperspectral feature of wheat biomass, several algorithms for sensitive variable selection were compared to determine the spectral variables for estimation model of wheat biomass. Synergy interval partial least squares (SIPLS) [1] and successive projections algorithm (SPA) [2] were employed to eliminate useless variables from the full hyperspectral data. On that basis an approach was proposed by combing SIPLS with SPA to determine the optimal spectra. Then, the optimal features were considered as input variables of the partial least-squares regression (PLSR) method [3],which was the mostly used calibration model and regression method. The determination coefficient of calibration (R ² _C ), the root mean square error (RMSE _V ), relative root mean square error of validation (RMSE _V ) and the number of input variables were presented to compare the performance of different methods in extracting sensitive spectral information.