Leaf water content estimation by functional linear regression of field spectroscopy data

Abstract

Grapevine water status is critical as it affects fruit quality and yield. We assessed the potential of field hyperspectral data in estimating leaf water content (Cw) (expressed as equivalent water thickness) in four commercial vineyards of Vitis vinifera L. reflecting four grape varieties (Mencia, Cabernet Sauvignon, Merlot and Tempranillo). Two regression models were evaluated and compared: ordinary least squares regression (OLSR) and functional linear regression (FLR). OLSR was used to fit Cw and vegetation indices, whereas FLR considered reflectance in four spectral ranges centred at the 960, 1190, 1465 and 2035 nm wavelengths. The best parameters for the FLR model were determined using cross-validation. Both models were compared using the coefficient of determination (R2) and percentage root mean squared error (%RMSE). FLR using continuous stretches of the spectrum as input produced more suitable Cw models than the vegetation indices, considering both the fit and degree of adjustment and the interpretation of the model. The best model was obtained using FLR in the range centred at 1465 nm (R2 = 0.70 and %RMSE = 8.485). The results depended on grape variety but also suggested that leaf Cw can be predicted on the basis of spectral signature.