Coupling hyperspectral data with principle component regression (PCR) and partial least square regression (PLSR) to improve prediction accuracy of rice crop variables

Abstract

Hyperspectral canopy reflectance data were used for principle component regression (PCR) and partial least squares regression (PLSR) model to predict two crop variables: leaf area index (LAI) and SPAD value for 3 cultivars of rice, 4 levels of nitrogen supply in the northern Java, Indonesia. Coefficient of determination (R²), cross validated determination of coefficient (R²CV), root mean square error in prediction (RMSEP), root mean square error of cross validation (RMSECV), root mean square error entire calibration (RMSEC) of model calibration and validation were calculated for the model quality evaluation. In the present study the effectiveness and practicability of multivariate analysis methods of PCR and PLSR compared and tested over all single wavebands of reflectance and first derivative reflectance (FDR). The predictive capability of PLSR model demonstrated slightly better predictive capability than that of PCR model attributed to LAI and SPAD, indicated by rising of R² and reducing the number of latent variable (NLV) and RMSEC. The predictability (R²) of crop variables was improved using PLSR, particularly in LAI with employing single reflectance (R² = 0.956) and FDR (R²=0.956). Using PCR and PLSR improved the accuracy (R²) model when compared to using simple linear regression attributed to predict LAI and SPAD value which we have done previously. Validation of measured and predicted value using PLSR model implied better accuracy to predict LAI and SPAD than that of using PCR model.