Oilseed rape is an important source of edible oil,zinc(Zn)is an essential element in the growth of oilseed rape,and the right amount of zinc fertiliser can promote the synthesis of chlorophyll in oilseed rape and accelerate its growth.However,the current situation of Zn pollution in soil is not optimistic,and high concentration of Zn pollution can inhibit the growth of oilseed rape,affect the yield and even endanger human health,so it is necessary to test the Zn content in oilseed rape.Traditional methods of Zn detection are tedious and time-consuming.Hyperspectral image technology is an emerging non-destructive testing technology,which has the advantage of image and spectrum,and the non-destructive detection of Zn content in rape leaves can be achieved without complicated operations by the hyperspectral images.However,it has the problems of large data dimension,redundant information and strong correlation between bands,which requires feature extraction to eliminate redundant information and simplify the model.Stacked Sparse Auto-Encoder(SSAE)is widely used in feature extraction of high-dimensional data because of its powerful feature learning capability.However,the network structure parameters of traditional SSAE are determined by manual experience,which is subject to chance and prone to overfitting when dealing with small samples,high-dimensional and nonlinear datasets,which is not conducive to the extraction of hyperspectral image features.Therefore,this paper proposed a Modified Stacked Sparse Auto-Encoder(MSSAE)to extract the deep fusion features of hyperspectral images of oilseed rape leaves,and built a prediction model based on the features extracted by MSSAE to achieve accurate detection of Zn content in oilseed rape leaves,and the main research contents and conclusions were as follows:(1)Cultivation of oilseed rape and acquisition of hyperspectral images.Oilseed rape with different Zn contents were obtained from oilseed rape stress experiments using 6 different concentrations of Zn solutions,and the true Zn content of the leaves was measured using flame atomic absorption spectrometry.The hyperspectral imaging system was used to acquire images of oilseed rape leaves under different concentrations of Zn solution stress.The whole leaf was extracted as the region of interest by threshold segmentation method,and the average spectral value of the region of interest was calculated as the raw spectral data,in addition,24 color features and 10 texture features of oilseed rape leaves were extracted.(2)Improvement of SSAE algorithm.Although the emergence of SSAE has greatly improved the feature learning capability of autoencoder algorithm,there were still problems such as network parameters relying on empirical settings and easy overfitting.In this paper,we solved the problem of overfitting in feature extraction by introducing Dropout mechanism instead of traditional KL scattering for sparse constraint,and optimized the network structure of SSAE by GWO to select the best number of hidden layer layers and neurons.The modified SSAE was used in feature extraction of hyperspectral images of oilseed rape leaves and compared with the traditional method.The results found that MSSAE is more suitable for extracting deep features of hyperspectral image data.(3)A model for prediction of Zn content in oilseed rape leaves based on spectral information was developed.The spectral curves were first preprocessed using Multiplicative Scatter Correction(MSC)to reduce the scattering and baseline drift of the curves.The preprocessed spectral data were extracted using MSSAE for feature extraction,and the feature extraction results were compared with the Variable Iterative Space Shrinkage Approach(VISSA)combined with the Competitive Adaptive Reweighed Sampling(CARS)algorithm and Successive Projections Algorithm(SPA).Finally,a Least Squares Support Vector Regression(LSSVR)prediction model was built based on the full-spectrum information and the features extracted by different algorithms,and the parameters of the model were optimized using the Cuckoo Search(CS)algorithm.The results showed that the model based on MSSAE extracted features had the best prediction results,with2((8) of 0.958,of 1.098 mg/kg,2of 0.940 andof 1.245 mg/kg.The results indicated that the use of MSSAE extracted features based on spectral information was able to detect Zn content in oilseed rape leaves.(4)A model for prediction of Zn content in oilseed rape leaves based on the graph-spectrum fusion information was developed.To further improve the model accuracy,the spectral information,color features and texture features of rape leaves were normalized and fused,and different algorithms were used to extract features from the graph-spectrum fusion information and build the prediction model.The comparison revealed that the best prediction model was built based on 28 features extracted by MSSAE.After that,the model was optimized using CS,and it was found that the MSSAE-CS-LSSVR model had the highest accuracy(2(8) of 0.979,of 0.931 mg/kg,2of 0.951 andof 1.127 mg/kg.Therefore,the MSSAE-CS-LSSVR model based on the graph-spectrum fusion information could effectively improve the detection accuracy of the model,which also further indicated that the deep learning method combined with hyperspectral images could achieve accurate and nondestructive detection of Zn content in oilseed rape leaves.In this paper,we proposed the MSSAE based on SSAE,and used the MSSAE to extract the depth features of the hyperspectral images of rape leaves and establish a prediction model,which successfully achieved the high precision and nondestructive detection of Zn content in rape leaves.This provided technical support for the rapid and nondestructive detection of Zn content in oilseed rape leaves,and also provided a reference for the nondestructive detection of heavy metal content in other plants. |