Analysis and Detection Method of NIR Spectral Characteristics of Kidney Bean Canopy Under Saline-Alkali Stress
WANG Lu1, GUAN Hai-ou1*, LI Wei-kai2, ZHANG Zhi-chao1, ZHENG Ming1, YU Song3, HOU Yu-long3
1. College of Electrical and Information, Heilongjiang Bayi Agricultural University, Daqing 163319, China
2. Northeast Agricultural University, Harbin 150030, China
3. College of Agriculture, Heilongjiang Bayi Agricultural University, Daqing 163319, China
Abstract：Salinity-alkalinity stress is one of the important of adversity factors that affect the kidney bean production and quality, the research of crop salinity-alkalinity stress is commonly by conventional chemical milling extraction method, the operation is complicated and time-consuming and destructive, such as for kidney bean canopy under salinity stress near-infrared (NIR) spectrum feature extraction, and quick nondestructive testing its salinity-alkalinity stress degree research rarely reported. In order to solve the problem of rapid detection of salt and alkali stress of kidney bean at the seedling stage, a new method for analyzing and detecting the NIR spectral characteristics of the canopy of kidney bean under salt and alkali stress was proposed based on near-infrared spectroscopy to study the characteristics of healthy and multi-grade salt and alkali stress of kidney bean at the seedling stage. Firstly, the spectral data of kidney bean canopy with healthy seeding stage and saline-alkali stress in the range of 990~2 452 nm with strong absorbance value were selected for study, and the original spectral data were pre-processed by using the automatic fitting detrend algorithm (DT) with a quadratic polynomial. Then a competitive adaptive reweighted sampling algorithm (CARS) was selected to extract 95 characteristic wavelengths sensitive to saline-alkali stress from the pre-processed data. The radial basis function was used as the hidden neuron to construct a three-layer feedforward neural network structure of type 95-282-7 (RBF). The network parameters were determined through the training set of samples, and the forward output value of the network was coded as a binary vector. Finally, the output vector was analyzed to the saline-alkali stress degree and the rapid detection method of saline-alkali stress degree of a kidney bean at the seeding stage was completed. The results showed that: (1) the original spectral curve was preprocessed in a variety of ways, and the correlation range of the study results was 0.339 4~0.946 1, the correlation range of DT pretreatment spectrum was 0.943 3~0.946 1, and the mean value was only 0.944 7, which could improve the accuracy of rapid detection of salt and alkali stress of kidney bean. (2) aiming at the near-infrared spectrum curve of kidney bean canopy pretreated by DT, CARS algorithm was optimized to extract the spectral characteristic wavelength vector of 95 dimensions. The total wavelength of kidney bean was reduced by 93.51%, effectively preserving the characteristic information source sensitive to salt and alkali stress. (3) application of CARS - RBF model for automatic rapid detection of kidney bean salinity-alkalinity stress degree in the study of 282 times, the mean square error (MSE) is 0.009 938 59, model checking accuracy reached 97.73%, so, this method is a new way of rapid non-destructive detection of saline-alkali stress degree of kidney bean, and can provide a technical reference for the rapid non-destructive detection of the saline-alkali stress degree of other crops.
