Objective  To realize the rapid and nondestructive identification of rice sheath blight at the early stage using hyperspectral imaging technology, and establish the corresponding identification model.

Method  Healthy rice seedlings and rice seedlings infected with sheath blight were used as the research samples. A total of 360 hyperspectral curves of 180 samples were collected separately from leaves and canopy at the wavelength of 380 to 1 030 nm. After eliminating the obvious noise part, the spectra of rice samples were reserved at the wavelength of 440 to 943 nm, and the spectral curves were preprocessed with different treatments. Partial least squares-discriminant analysis (PLS-DA) was used to model different preprocessed spectra. The feature information was extracted from the original canopy spectral data using the MNF algorithm, and a linear discriminant analysis (LDA) model and a back-propagation neural net-work (BPNN) discriminant model were established based on the feature information.

Result  The prediction accuracy of PLS-DA model after preprocessing by standard normal variate transformation (SNV) was the highest (92.1%). The discriminant results of LAD and BPNN model based on feature information were superior to the PLS-DA discriminant model based on all bands. The BPNN model based on feature information from minimum noise fraction transformation had optimal results. The accuracies of the model set and prediction set were 99.1% and 98.4% respectively.

Conclusion  It is feasible to identify nondestructively rice sheath blight using hyperspectral imaging technology. Our research provides a new method for identifying rice sheath blight.