| Maize(Zea mays L.)is one of the most widely cultivated crops in the world,used for feeding,industrial processing,and food.Seed quality is the key factor to determine whether seeds can germinate and grow into seedlings.In the process of planting,harvesting,drying,transportation and storage,the quality of seeds will change due to the influence of external environmental conditions and their life activities,thus affecting the level of agricultural yield.In this study,near-infrared hyperspectral imaging technology was used to obtain the phenotypic information of crop seeds.Nondestructive detection was carried out on four indicators(damage state,variety attribute,ear rot infection,mildew infection)that affect the quality of crop seeds during seed processing.Based on the machine learning methods,models for crop seed damage recognition,variety identification,ear rot diagnosis and mildew detection were established.The main research contents and results are as follows:(1)For mechanical damage of seeds reducing seed clarity,an identification method of seed damage based on the fusion of spectral and textural features was proposed,and a nondestructive recognition model of maize seed damage state based on hyperspectral imaging technology was established.Hyperspectral images of intact and damaged seeds of various varieties were acquired,and spectral features were extracted using the successive projection algorithm(SPA).Then the textural features were extracted by combination with the gray-level co-occurrence matrix(GLCM).The convolutional neural network-bidirectional long short-term memory(CNN-Bi LSTM)model based on the fusion features achieved the optimal recognition effect.The accuracy of the training set,validation set and testing set was 99.31%,99.13% and 98.61%,respectively.Transfer component analysis(TCA)was used to transform the fusion features of different seed varieties,and the testing accuracy was 98.81%,indicating that the model had strong transferability.(2)For seed variety mixture reducing seed purity,the influence of training set scale on different convolutional neural networks was explored,and a nondestructive identification model of crop seed varieties based on hyperspectral imaging technique was established.Hyperspectral images of maize seeds of several varieties were acquired,and CNN models were established on different scales of training sets.It was found that the accuracy of the validation set and testing set of the different models fluctuated with the increase of the training set scale.The Res Net model based on the entire training set achieves the optimal identification effect,and the accuracy of the training set,validation set and testing set was 100%,98.36% and 98.20%,respectively.For the optimal Res Net model,the t-distributed stochastic neighbor embedding(TSNE)algorithm was used to visualize how the model gradually learned the differences of spectral features among different varieties.The visual identification of samples in the external testing set was realized,and the accuracy reached 98.00%.Based on the fine-tune method,a variety identification model within the class(maize seeds)was established.The accuracy of the training set,validation set and testing set was 99.25%,95.13% and 95.37%,respectively.In addition,a variety identification model between the class(grape seeds)was established,and the accuracy of the training set,validation set and testing set was 98.75%,97.00% and 96.78%,respectively,indicating that the model had strong generalization ability.(3)For ear rot infection affecting seed quality,the spectral characteristics of seed endosperm side and embryo side were explored,and a nondestructive diagnosis model of ear rot disease of maize seeds based on hyperspectral imaging technique was established.Hyperspectral images of the endosperm side and embryo side of healthy seeds,mildly infected seeds and severely infected seeds were acquired,and 1D-CNN and 2D-CNN models were established,respectively.For the full spectra,the diagnostic effect of the 1D-CNN model based on the endosperm side and embryo side was similar.The diagnostic effect of the 2D-CNN model based on the endosperm side was better than that of the embryo side.Based on band ratio operation and correlation analysis,15 and 27 characteristic indexes were generated for the endosperm and embryo sides,respectively.Then,characteristic images were generated based on characteristic indexes.The 1D-CNN models based on characteristic indexes and 2D-CNN models based on characteristic images were established,respectively.For the endosperm side and embryo side,the diagnostic effect of the 1D-CNN model was similar to that based on full spectra.In contrast,the diagnostic effect of the 2D-CNN model was superior to that based on full spectra.Especially for the embryo side,the negative influence of the folded area of the embryo part on the diagnosis effect of the model was greatly weakened,and the diagnosis effect of the2D-CNN model was greatly improved.In general,the 1D-CNN and 2D-CNN models established based on the full spectra,characteristic indexes and characteristic images of the endosperm side achieved good diagnostic effect.Thus,the endosperm side could be used as the dominant side for the diagnosis of ear rot.The testing accuracy of the 2D-CNN model based on the characteristic images of the endosperm side reached 91.67%.(4)For mildew infection reducing seed vigor and storability,the spectral characteristics of seeds with different mildew degrees and the saliency map analysis method of different dimensional models were studied,and a nondestructive detection model of mildew state of maize seeds based on hyperspectral imaging technique was established.Hyperspectral images of healthy seeds,mildly mildewed seeds,moderately mildewed seeds and severely mildewed seeds were acquired,and 1D-CNN and 2D-CNN models were established.For detecting whether the seeds were mildewed or not,the 1D-CNN model had a better detection effect,with the accuracy of 98.75%,98.33% and 95.00% in the training set,validation set and testing set,respectively.Combined with the saliency maps of 1D-CNN and 2D-CNN models,it was found that in the 2D-CNN saliency maps corresponding to 985.08,1011.93,1015.29,1045.52,1048.88,1069.03,1079.11 and 1102.64 nm,the gradient updating direction of most pixels in healthy seeds and mildewed seeds was opposite.For detecting healthy seeds and seeds contaminated with different degrees of mildew,the 1D-CNN model had a better detection effect,with the accuracy of 100%,96.67% and 95.00% in the training set,validation set and testing set,respectively.Combined with the saliency maps of 1D-CNN and 2D-CNN,it was found that in the 2D-CNN saliency maps corresponding to 1015.29,1018.65,1022.01,1106.00,1119.45,1126.17,1217.02,1307.97,1375.42,1635.66,1642.43,1645.82 nm,as the seeds changed from health state to increasing degrees of mildew,the gradient updating direction of most pixels showed a consistent trend(positive increase at some wavelengths and negative decrease at others). |
PDF Full Text Request