Inversion of soil Cu concentration based on band selection of hyperspetral data

Hyperspectral data offers a powerful tool for predicting soil heavy metal contamination due to its high spectral resolution and many continuous bands. Band selection, however, is the prerequisite for heavy metal inversion by hyperspectral data. In this study, soil reflectance spectra and their Cu contents were measured for 181 soil samples in the laboratory. Based on these dataset, band selection was conducted to inverse Cu contents using stepwise regression approach, and prediction accuracies of Cu based on partial least-squares regression (PLSR) model with different selected bands were analyzed. In addition, the influences of spectral resolution on prediction results of Cu were discussed by a Gaussian re-sampling function. It demonstrated that the optimal band number was 10 for Cu inversion and the corresponding model had prediction accuracy of R2 = 0.7523 and RMSE = 0.4699; the optimal spectral resolution was 32nm and the model on this basis had an accuracy of R2 =0.7028 and RMSE =0.5147. Results of this study may provide scientific verification for designing low-cost and practical hyperspectral spaceborne sensors, and theoretical bases for simulating spaceborne sensors to predict soil heavy metals contents in the future.