Statistical models for prediction of dry weight and nitrogen accumulation based on visible and near-infrared hyper-spectral reflectance of rice canopies

Much information is obtainable from hyper-spectral data, which measure solar radiation consecutively at less than about 10-nm intervals. In constructing statistical prediction models, however, problems of overfitting may arise due to the excessive number of variables, and multicollinearity may occur between variables ; thus a few specific wavelengths must be chosen. Various multivariate regression models were examined with ten-fold cross-validation to develop efficient, accurate models to predict dry weight and nitrogen accumulation of rice crops from the maximum tiller number stage to the meiosis stage, using plant-canopy reflectance of hyper-spectra within the 400-1100 nm domain without any variable selection. The results showed that the principal component regression using hyperspectra gave better fits and predictability than that using specific wavelengths. On the other hand, partial least squares regression was the most useful among the models tested ; this method avoided overfitting andmulticollinearity by using all wavelength information without variable selection and by inclusion of both x and y variations in its latent variables.