[Estimating the severity of rice brown spot disease based on principal component analysis and radial basis function neural network].

An ASD Field Spec Pro Full Range spectrometer was used here to acquire the spectral reflectance of healthy and disease leaves cut from rice plants in the field. The leaf disease severity of rice brown spot was determined by estimating the percentage of infected surface area of rice leaves in the laboratory through phytopathologist's observation. Three steps were taken to estimate leaf disease severity of rice brown spot. The first step was that different spectra transforming methods, namely, resampling spectrum (10 nm interval), the first- and second-order derivative spectrum based on raw hyperspectral reflectance, were conducted. The second step was that the principal component analysis (PCA) was examined to obtain the principal components (PCs) from the above transformed spectra to reduce the spectra dimensions of hyperspectral reflectance and simplify the data structure of hyperspectra. The last step was that the resampling and PCs spectra entered the Radial Basis Function neural network (RBFN) as the input vectors, and the disease severity of rice brown spot entered RBFN as the target vectors. RBFN is an effective feed forward propagation neural network, which is based on the linear combinations of corresponding radial basis functions. In general RBFN can be used to solve the problems such as regression or classification with high operation rate and efficient extrapolation capability, and quickly designed with zero error to approximate functions. The total dataset (n = 262) was divided into two subsets, in which three quarters (n = 210) was the training subset to train the neural network, and the remaining quarter (n = 52) was the testing dataset to conduct the performance analysis of neural network. The spread constants of RBFN and various data processing methods were investigated in detail. The best prediction result was obtained by PCs spectra based on the first-order derivative using RBFN model, the root mean square of prediction error (RMSE) was small (7.73%) in the testing dataset, and the next was the resampling spectra with RMSE of 8.75%. This research demonstrated that it was feasible and reliable to estimate the disease severity of rice brown spot based on PCA-RBFN and hyperspectral reflectance at the leaf level.