Rice Varieties Identification And Amylase Activity Detection Based On Image Spectra Features And Multivariate Analysis Methods

High-quality rice has higher nutritional value and economic value.For the sake of high returns,some rice adulterated and shoddy rice are constantly appearing.These phenomena will damage purchasing rights of consumers,destroy the rice export trade market and discourage the enthusiasm of high-quality rice producers.To solve these problems,the rapid and accurate identification of high-quality rice is crucial.At the same time,storage is very important for rice that is waiting to be processed into rice or seed.The stored rice contains a certain amount of amylase,and the seed rice has a higher amylase activity after germination.Based on these differences,it is explored to determine the storage status of rice by accurately measuring the amylase activity value of different rice.Therefore,the main research of this article is to use image spectra features and multivariate analysis methods to identify rice varieties and detect rice amylase activity.The main research contents are as follows:(1)The ten types of high-quality rice were identified using deep learning methods and hyperspectral imaging(HSI).Principal component analysis network(PCANet)was applied to integrate the above characteristics to establish a classification model of rice types,and machine learning methods such as K nearest neighbors(KNN)and random forest(RF)were used for comparison.Meanwhile,Multiple pre-processing methods were applied to process the spectra,and principal component analysis(PCA)was performed to obtain the main information of high-dimensional features.The results showed that multi-feature fusion improved the recognition accuracy,and PCANet demonstrated considerable advantages in classification performance.The best result was achieved by PCANet with PCA-processed spectra and texture features with correct classification rates of 98.57%for prediction sets,respectively.The moisture content of rice was studied to help improve the distinction of rice varieties.The partial least squares regression(PLSR)model and spectral features were employed to analyze and realize rapid and non-destructive identification of rice moisture.Multiple pre-processing methods were used to process spectral features,and important wavelengths were selected based on the competitive adaptive re-weighting algorithm(CARS).The best model came from the raw full spectra and PLSR(R_p~2=0.7979,RMSEP=0.4848)..The results of moisture measurement were poor and failed to effectively help improve the identification of rice varieties.In conclusion,the varieties of rice were accurately identified based on the image spectra features of rice and the PCANet classification model.(2)The 400-2500 nm visible/near infrared reflectance spectra and hyperspectral images of four kinds of rice were collected,and combined with PLSR and support vector machine regression(SVR)to predict the amylase activity in the rice.Four methods of MSC,SNV,normalization and S-G smoothing were selected to preprocess the raw spectra,and CARS was adopted to select important wavelengths for the raw spectra and the preprocessed spectra.Experiments showed that the prediction model established by important wavelengths was better than full-spectra modeling.In addition,the texture features of rice were extracted based on hyperspectral images and three texture extraction methods,and combined into the spectra of important wavelengths.It was found that the integration of a single texture feature does not improve the detection effect of amylase activity.The combination of gray level-gradient co-occurrence matrix(GLGCM)and discrete wavelet transform(DWT)improved the prediction effect,and combined the normalized spectra of important wavelengths and PLSR modeling to obtain the best prediction results(R_p~2=0.8639,RMSEP=0.1668).It could be seen that the spectral and texture features of rice can achieve accurate detection of amylase activity.(3)The analysis software for rice varieties identification and amylase activity detection was designed and developed to realize image feature extraction,spectral preprocessing,feature dimensionality reduction,variable selection and modeling analysis.Modeling analysis included establishing classification model and prediction model.The studies demonstrated that the combination of image spectra features and multivariate analysis methods can achieve accurate identification of rice varieties and identification of amylase activity.The fusion of multiple features could describe the properties of rice more accurately.The model analysis ability of deep learning was significantly better than machine learning,and it could be widely used in data analysis.The method used in this study could be extended to the identification of other substances and the detection of internal components.