Testing genetic algorithm as a tool to select relevant wavebands from field hyperspectral data for estimating pasture mass and quality in a mixed sown pasture using partial least squares regression

It is now accepted that partial least squares (PLS) regression with waveband selection might improve their predictive accuracy in multivariate calibration of models to describe pasture mass and quality. Recently, genetic algorithm (GA) has been shown to be a suitable method for selecting wavebands in the laboratory calibrations. This study aimed to investigate the performance of genetic algorithms PLS (GA-PLS) regression analyses for estimating green forage biomass (GBM), neutral detergent fiber (NDF), acid detergent fiber (ADF), acid detergent lignin (ADL) and crude protein (CP) concentrations of herbage and herbage mass in CP (CPm) from field hyperspectral data at canopy scale. The predictive ability of GA-PLS was compared with that of iterative stepwise elimination PLS (ISE-PLS) and standard full-spectrum PLS (FS-PLS), using first derivative reflectance (FDR) spectra data. Canopy reflectance measurements and plant sampling were obtained from 50 selected sites in each of two seasons; spring (May) and summer (July) 2007. For all parameters, cross-validated coefficients of determination (R2) increased and root mean square error values decreased, respectively, with GA wavebands selection. The number of wavebands selected in the GA-PLS model ranged between 63 (4.0% of all 1563 available wavebands) and 167 (10.7%), suggesting that over 89% of wavebands were redundant. Between ISE-PLS and GA-PLS models, higher R2 values and lower root mean squared errors of prediction were obtained from GA-PLS prediction for all parameters except CP concentration. Particularly, GA based waveband selection greatly improved ADF (R2 = 0.68-0.77) and ADL (R2 = 0.47-0.59) predictions. These results suggest that pasture quality and GBM can be predicted from field hyperspectral measurements using a GA-PLS model, and that the GA-PLS has the advantage of tuning the optimum bands for PLS regression, giving better predictive ability.