Estimating forage biomass and quality in a mixed sown pasture based on partial least squares regression with waveband selection

Although partial least squares (PLS) regression, a full-spectrum bilinear regression method, is widely used in laboratory calibrations of pasture quality, increasing evidence indicates that PLS models include some redundant wavelengths. Consequently, more careful wavelength selection might improve their predictive accuracy, especially in field applications. We compared the predictive ability of PLS models using whole and selected wavebands from in situ canopy reflectance spectra over 400-2350 nm to predict above-ground biomass (BM) and concentrations of crude protein (CP), acid detergent fiber (ADF) and neutral detergent fiber in herbage. Canopy reflectance measurements and plant sampling were conducted at 86 selected points in a mixed sown pasture in Hokkaido, Japan, in August 2006. Removing the minimum value of the weighted regression coefficient in the PLS model enabled stepwise waveband selection. For all pasture parameters, cross-validated coefficients of determination ( [graphic removed] ) and root mean square error values, respectively, increased and decreased with removal of wavebands until the optimum number of wavebands was reached. The number of selected wavebands ranged between six (2.2% of full 277 wavebands) and 47 (17%), suggesting that over 83% wavebands were redundant or useless. Overall, higher R² values and lower root mean squared errors of prediction were obtained using selected wavebands in the PLS model. Particularly, waveband selection greatly improved BM (R² = 0.51-0.72) and ADF (R² = 0.30-0.65) predictions when using the first derivative reflectance spectrum, and CP (R² = 0.38-0.62) prediction for reflectance. These results suggest that pasture quality and BM can be predicted by in situ canopy reflectance using a PLS regression model, and that the predictive ability of the model can be improved by optimizing important wavebands.