Abstract

Water deficit stress (WDS) can negatively affect the development, productivity and quality of Eucalyptus spp. To minimize this impact, the development of high throughput techniques for early and accurate WDS detection is necessary. This study focuses in the use of visible-near infrared hyperspectral imaging (VIS-NIR HSI) and chemometric methods to detect water deficit spectral patterns of these species and to develop predictive models for early detection of water deficit level in juvenile plants. The research included the analysis of four Eucalyptus genotypes, two Eucalyptus globulus and two Eucalyptus gloni (hybrid of E. nitens and E. globulus). Forty ramets of each genotype were submitted to WDS conditions associated with different stress levels and compared with control samples using conventional physical characterization and common vegetative indices of plants, besides VIS-NIR HSI. The HSI data were analyzed using principal component analysis (PCA) and supervised pattern recognition methods to classify the samples by WDS level using as validation set the mean spectra of images (bulk prediction) and all the pixels of whole plant (single pixel prediction). The results of PCA showed a differentiated response on the different WDS conditions, especially at day 10 of water deficit. Conventional vegetation indices, such as NDVI, PRI and MCARI, did not detect indications of an early WDS response, while pattern recognition methods including partial least squares (PLS-DA), discriminant support vector machine (SVM-DA) and k-nearest neighbor (KNN) showed a remarkable predictive ability for WDS level with a prediction error (Err) of 2 % in external validation sets. Supervised models were applied also to reconstruct the stress level in all the pixels of whole plants. The superior effectiveness of the SVM-DA and KNN models to predict stress level in images of Eucalyptus spp. provided valuable information on spatial distribution of stress in the plant.