Machine Learning- And Hyperspectral Imaging- Based Nondestructive Detection Approaches And Its Applications

As an emerging high-precision non-destructive testing technology,hyperspectral imaging is widely used in agricultural fields,such as wood quality detection,identification of plant leaf diseases,and quality detection of vegetable and fruits or other foods.The moisture content is an important attribute of wood,and also is an influence factor of the quality of wood.Plant leaf disease is a prevalent and severely harmful to the growth of plants,and directly affects the leaves’photosynthesis efficiency.Due to many spectral bands and high redundancy,it is necessary to reduce the data’s dimensionality.This paper takes wood and loquat leaves as the research object,studies and improves the commonly used spectral data dimensionality reduction methods,applies machine learning and hyperspectral imaging to determine the moisture content of wood,and classifies loquat leaves with different disease levels.The main research contents are as follows:1.Aiming at the rapid non-destructive testing of wood moisture content,a wood moisture content measurement method based on modified random frog(MRF)and Gaussian process regression(GPR)is proposed.Since the random frog needs to reset threshold after obtaining the selection probability of all variables,adaptive reweighted sampling(ARS)and exponential decreasing function(EDF)are applied to improve RF.A total of 30 beech blocks were used in the experiments,each block was dried seven times to generate different cases of moisture contents.The 210 samples were divided into a calibration set(140)and a prediction set(70)using Kennard-stone(KS)algorithm.The spectral data of each sample was collected by the hyperspectral imaging system,and pretreated using standard normal variate(SNV).A Gaussian process regression model was established based on four feature selection algorithms:modified random frog,random frog,successive projections algorithm(SPA),and competitive adaptive reweighted sampling(CARS).The results show that the MRF not only avoids setting thresholds but also improves the stability and accuracy of the model.The MRF-GPR model achieved the best predictive performance of wood moisture content,with Rp2,RMSEP of0.9785,1.6125%,respectively.2.Aiming at the problem that the iterative retained information variable(IRIV)is easy to fall into the local optimum,a modified iterative retained information variable(MIRIV)is proposed.Based on the original algorithm,MIRIV adopts sequential floating forward selection(SFFS)algorithm to select the strongly and weakly informative variables retained after iteration,effectively solving the problem of backward elimination algorithm that is easy to fall into local optimality.To fully verify the feasibility of MIRIV,this article uses two public datasets and one private dataset.On these three datasets,competitive adaptive reweighted sampling(CARS),random frog(RF),iteratively retaining informative variables(IRIV),and MIRIV are used to filter feature variables,and the partial least squares(PLS)prediction models are established based on different selection methods.The experimental results show that,in most cases,MIRIV has the best dimensionality reduction performance.For example,in the wood moisture content dataset,MIRIV selected only 15 characteristic bands among 128 bands,and the R_p~2and RMSEP of the model are 0.9569 and 2.1432%,respectively.3.Aiming at rapid identification of loquat leaves with different disease levels,a leaf classification method based on linear discriminant analysis(LDA)and hyperspectral imaging technology is proposed.LDA,K nearest neighbor(KNN),and support vector machine(SVM)optimized by cuckoo search(CS)are used as classification models.On the campus of Jiangnan University,20 pieces of four loquat leaves with different disease levels were picked,and the four types of loquat leaves were:mature and healthy,immature and healthy,mild infection,severe infection.After dust removal and drying,all the leaves were photographed with a hyperspectral imager,and ten regions of interest were selected on the surface of the leaves through ENVI5.3.A total of 800 spectral data were obtained.Savitzky-Golay smoothing is used to preprocess the spectral data,and 800 samples are divided into training set(560)and prediction set(240)using the Kennard-stone(KS).Three classification models are established based on the full-spectrum band.Experimental results show that the LDA classification model has the highest recognition accuracy,with a classification accuracy rate of 94.6%,and a total of 227 samples are recognized accurately.