Abstract:

Hyperspectral techniques have been widely used to monitor the degree of drought stress and screen for drought-resistant plant materials. In this study， 18 bermudagrass genotypes were subjected to natural drought for 12 days. The soil water content and leaf relative water content of each material were determined， and hyperspectral photos were obtained at different timepoints during the drought treatment using an SVC HR-768 portable spectrometer. The spectral reflectance， which was processed using a combination of hyperspectral Savitzkye Golay smoothing and Savitzkye Golay smoothing+first derivative， was used as the independent variable， and characteristic bands that were well correlated with the relative water content of leaves and， on the basis of Pearson’s correlation and continuous projection analyses， were common among all the materials screened. Then， an inversion model of the relative water content of bermudagrass leaves was established using three machine-learning algorithms： the BP neural network， support vector machine， and random forest algorithms. The main results were as follows： 1） Five Savitzkye Golay smoothing+first derivative characteristic bands were screened out using the continuous projection algorithm； the bands were located at 406， 569， 706， 736， and 786 nm， respectively， and showed high correlations （P>0.5） with the relative water content of bermudagrass leaves. The covariance among the bands was weak， so these bands could be used as sensitive bands for drought monitoring. 2） The coefficient of determination （R² ） and root-mean-square error （RMSE） of the random forest inversion model based on the sensitive wavebands were 0.939 and 8.552， respectively， which were 5% and 8% higher， respectively， than those of the support vector machine and BP neural network models. Thus， the random forest inversion model showed the best prediction effect and universality. The R² of the test set was 0.925， and the RMSE was 9.008. The results of our study provide technical support for the development of a non-destructive and high-precision method to monitor the relative water content of bermudagrass leaves using hyperspectroscopy.

Key words: bermudagrass, different genotypes, leaf relative water content, hyperspectral techniques