Abstract: The incidence of disease and insect stresses in rice panicles at the late growth stage had a negative impact on normal growth of panicles. Discrimination of health conditions in stressed rice panicles from healthy ones was the basis of plant protection measurements and yield assessment. Five kinds of health condition in rice panicles were classified in the study： no infection, light and moderate infection caused by rice glume blight disease, serious infection caused by rice false smut disease, and white headed panicle caused by either the stem borers Chilo suppressalis or rice panicle blast. The wavelength range of hyperspectral reflectance of rice panicles was measured from 350 to 2 500 nm with a portable spectroradiometer in laboratory conditions. The spectral response characteristics of rice panicles were investigated, and principal component analysis (PCA) was performed to obtain the principal components (PCs) derived from different transforming spectra, namely raw (R), inverse logarithmic (Ln 1/R), first, and second derivative reflectance (FDR and SDR). PCs was input into the learning vector quantization (LVQ) neural network classifier as the input vectors to classify healthy, light, moderate, and serious infection levels and white headed panicles. Classification accuracy was evaluated using overall accuracy and Kappa coefficient. The overall accuracies of LVQ with PCA derived from R, Ln 1/R, FDR and SDR for the validation dataset were 75.3%, 74.7%, 91.6% and 100% respectively, and the corresponding Kappa coefficients were 0.689, 0.682, 0.895 and 1.000. The results demonstrated that it is feasible to discriminate different health conditions of rice panicles under laboratory conditions using hyperspectral remote sensing．

Key words: rice, hyperspectral remote sensing, principal component analysis, artificial neural network