Abstract: In order to realize rice nitrogen nutrition diagnosis quickly and accurately, rice cultivation experiments with 4 nitrogen application levels were conducted on Zhongjiazao 17 rice cultivar. A total of 240 rice spectral data were collected at tillering stage within 350 to 2 500 nm from top trilobate by using portable earth mass spectrometry. All the samples were randomly divided into training set (160 samples) and prediction set (80 samples). The original spectrum was pretreated by multiplicative scatter correction (MSC), standard normal variate (SNV) and Savitzky-Golay smoothing (SG) methods, respectively. Then, principal component analysis (PCA) and successive projection algorithm (SPA) were used for feature reduction and feature selection of the pre-processed spectra. After principal component analysis, the first 24 principal components, of which the accumulative contribution exceeded 99.98%, were used as the input variables of the model, and 12, 15 and 19 characteristic wavelengths were selected for the spectrum after MSC, SNV and SG treatments, respectively. Finally, rice nitrogen nutrition diagnosis models were established with support vector machine (SVM). It was shown that the MSC-PCA-SVM model was the best method for rice nitrogen nutrition diagnosis, of which the accuracy rate on training set and prediction set was 99.38% and 97.50%, respectively.

Key words: rice leaf, nitrogen nutrition diagnosis, hyperspectrum, principal component analysis, successive projection algorithm, support vector machine