Rice Origin Identification Based On Hyper-spectral Imaging Technology

In order to explore a fast,accurate and stable method for confirming the rice origins based on hyper-spectral image technology that 990 rice samples are collected from different rice producing areas in Jilin Province as research objects.Hyper-spectral imaging systems are used to obtain hyper-spectral images in the 400~1000nm range,and the average spectral reflection information in the 10pixel×10pixel region of interest was extracted as sample data.In order to reduce the effect of noise and other interference information,standard normal transformation(SNV),multiplicative scatter correction(MSC)and convolution smoothing(Savitzky-Golay,S-G)are used to pretreat the spectral curve.Three no-linear machine learning algorithms include Multi-Layer Perceptron(MLP),Extreme Learning Machine(ELM)and Online Sequential Extreme Learning Machine(OS-ELM)and Partial Least Squares Regression(PLS)are used to calculate during this test;the origin confirmation model is established based on the spectral data of the full-band and the spectral data after Adaptive Re-weighted Sampling(CARS),Principal Component Analysis(PCA)and Multidimensional scaling(MDS)dimension reduction processing,compares and analyzes the classification accuracy and training time of the model.Conclusions:(1)During the confirmation model based on the full-band spectral data,the average accuracy of the OS-ELM model is higher than ELM and MLP models whether the spectral curve is preprocessed or not.The MSC method is superior to SNV and SG in the pre-processing results,the classification accuracy of the MSC-OS-ELM model is the highest(98.3%),follows with MSC-ELM(89.6%),MSC-MLP(88.4%)and MSC-PLS(79.5%).(2)15 features bands are extracted with CARS dimension reduction processing,13 feature bands are extracted with MDS dimension reduction processing and 9 feature bands are extracted with PCA dimension reduction processing are used to analyze classification accuracy.The results show that the highest is MDS-OS-ELM(97.4%),then follow with MDS-ELM(88.5%),MDS-MLP(87.1%)and MDS-PLS(78.3%).The classification accuracy of the all four models decrease,but the input variables decrease 96.6%;Under the premise of ensuring higher accuracy,the model running time is sharply reduced.(3)In order to test model training efficiency with same parameters,two conditions include data that inputs 500 groups for once and data that inputs for 5 times to 500 groups are compared;the training time of first condition in turn goes ELM(0.0301s),OS-ELM(0.0303s),MLP(246.5643s),PLS(436.6958s);the gap between ELM and OS-ELM is very small,and both are significantly better than the MLP model;MLP and PLS are removed in the second condition,just compares the training time between ELM and OS-ELM and finds the training time of ELM shows linear uptrend with accumulation of sample data,however the training time of OS-ELM doesn't change apparently.During the origin confirmation application,the truth is always inputting data for times,so OS-ELM algorithm is more suitable for rice origin confirmation.The consequences show that MDS dimension reduction method can effectively extract the spectral information of rice and high-spectral imaging technology combined with OS-ELM algorithm is feasible for the efficient,accurate and stable non-destructive identification of rice origin confirmation.