Comparative analysis of different uni- and multi-variate methods for estimation of vegetation water content using hyper-spectral measurements

Assessment of vegetation water content is critical for monitoring vegetation condition, detecting plant water stress, assessing the risk of forest fires and evaluating water status for irrigation. The main objective of this study was to investigate the performance of various mono- and multi-variate statistical methods for estimating vegetation water content (VWC) from hyper-spectral data. Hyper-spectral data is influenced by multi-collinearity because of a large number of (independent) spectral bands being modeled by a small number of (dependent) biophysical variables. Therefore, some full spectrum methods that are known to be suitable for analyzing multi-collinear data set were chosen. Canopy spectral reflectance was obtained with a GER 3700 spectro-radiometer (400–2400 nm) in a laboratory setting and VWC was measured by calculating wet/dry weight difference per unit of ground area (g/m²) of each plant canopy (n = 95). Three multivariate statistical methods were applied to estimate VWC: (1) partial least square regression, (2) artificial neural network and (3) principal component regression. They were selected to minimize the problem related to multi-collinearity. For comparison, uni-variate techniques including narrow band ratio water index (RWI), normalized difference water index (NDWI), second soil adjusted vegetation index (SAVI2) and transferred soil adjusted vegetation index (TSAVI) were applied. For each type of vegetation index, all two-band combinations were evaluated to determine the best band combination. Validation of the methods was based on the cross validation procedure and using three statistical indicators: R², RMSE and relative RMSE. The cross-validated results identified PLSR as the regression model providing the most accurate estimates of VWC among the various methods. The result revealed that this model is highly recommended for use with multi-collinear datasets ($R_{\text{CV}}^2=0.94$, RRMSE_CV = 0.23). Principal component regression exhibited the lowest accuracy among the multivariate models ($R_{\text{CV}}^2=0.78$, RRMSE_CV = 0.41).