Research On Wheat Phenotypic Information Perception Method Based On Spectrum And Image

Wheat is one of the three major food crops.Accurate and efficient information perception and management methods are significant to wheat cultivation and food security.Wheat phenotypic information perception plays a key role in breeding,crop nutrient status evaluation and agricultural product quality inspection.In the traditional agricultural production mode,the process of breeding and seed selection rely on manual work,with the disadvantages of low efficiency and easy fatigue,which cannot meet the requirements of automatic and intelligent information perception in the process of agricultural modernization.Physical and chemical traits monitoring and food composition analysis rely on destructive experiments,which is hard to achieve green and zero-emission.Computer vision and spectral imaging technologies have recently provided non-destructive solutions for wheat information perception.Artifitial intelligence algorithms can correlate the collected digital signals and the internal and external physiological indexes of wheat crops or the quality and safety level of wheat products,providing new development opportunities for wheat phenotyping.Based on spectral analysis,machine vision,and artificial intelligence,this thesis explored wheat information perception methods,including wheat seed phenotyping,leaf nutrient status evaluation,plant variety identification and wheat mixture powder composition analysis.The details are provided as follows:(1)Wheat seed phenotyping based on hyperspectral imaging.During the process of spectral scanning,the seed samples are touching,which makes the individual seeds unseperatabe,and the redundant information of spectral data decreases the efficiency of phenotypic analysis.Therefore,a seed image segmentation method,called Seed Net,for conglutination pixels removal and a convolution neural network feature selection algorithm CNN-FS for spectral feature selection were proposed.This study collected 30 categories of wheat seeds and obtained their sprectral information using a line-scanning spectrometer.The Seed Net algorithm used a semantic segmentation network to distinguish conglutination,contour and seed main area.Then,connected components extraction was performed on the segmented seed area to generate the segmentation result.The Seed Net used a relatively simple semantic segmentation method to realize the instance segmentation of wheat seed.The counting accuracy and segmentation accuracy exceeded the traditional watershed segmentation,U-Net model and Mask-RCNN model.The spectral curve of individual seed could be obtained after segmentation,which were used for variety identification.The CNN-FS method used a feature selection layer constructed using nonlinear activation weighting and sparse penalty constraints,to screen the important spectral bands for wheat variety classification and achieve spectral dimention reduction.The classifier based on the features selected by the CNN-FS achieved a classification accuracy of 90%in a 30-class wheat seed identification task,which was higher than the classifier trained using the features screened by existing feature selection methods.(2)The method and device for leaf SPAD and water content inspection.The color features from the captured RGB image of leaves and the partial least square regression method were adopted for predicting the SPAD value.The model traversed all pixels in the leaf area to generate the SPAD distribution map.The light source,sensing unit and controller were integrated into a portable SPAD sensor.It realized an RMSE of 1.1 SPAD units when testing on wheat leaves.The average time for processing one sample was lower than 5s,achieving non-contact inspection and leaf SPAD distribution visualization.Moreover,this study developed a leaf water content sensor using NIR spectral module.The water content evaluation model was constructed based on a convolutional neural network and information fusion.Two hundred samples were collected for experiment.Testing results showed satisfactory performance,with RMSE of 0.79 SPAD and R~2 of 0.94 and the average time for a single test was 3s,which achieved rapid and non-destructive water content detection.(3)Wheat plant variety identification based on canopy information.The canopy images of five varieties of wheat plants at different growth stages were collected by UAV and smartphone.A dataset including 6400 images with 800*800 resolution was built.Popular deep learning classifiers including Res Net and Vision Transformer were employed to identify wheat plant variety.The Grad-CAM algorithm was used to visualize the classification judgment basis of the model for explaining the rationality of the trained classifier.Results showed that the highest discrimination accuracy could reach 87%by the Res Net model.Specially,the identification accuracy of plants at heading stage reached 90.56%,but the identification rate at early growth stage was only 46.71%.In this study,NIR spectroscopy was further used to identify the plants in the early growth stage.The discrimination model was established based on PLSDA and 1D CNN.The classification accuracy was improved to 77.5%.(4)Wheat flour identification and wheat mixture powder analysis based on NIR spectrum.Twelve kinds of food powders were obtained and scanned by an NIR spreal sensor.An extensible powder classification model was designed based on the fusion of deep learning and metric learning.Firstly,the six-class classification model was trained by using the six kinds of powder information,and the cosine similarity matching was introduced to replace the linear classifier of the initial model.By adding a very small number of reference samples of six new categories to the feature matching library,the classification model could accurately identify the samples of new categories without re-training or fine tuning.The results showed that the accuracy of 12-class classification was 97.86%.Then,280 multi-mixture powders containing wheat flour,corn flour,rice flour and milk powder were prepared and their NIR spectra were collected.The analytical models based on PLSR and CNN multiple regression were studied for proportion determination of each component in the mixture.The results showed that CNN multiple regression model performed better than the traditional PLSR method.The mean values of R~2 and RMSE on the test set were 0.976%and 0.035 respectively.The NIR sensor unit,Raspberry Pi and the models were integrated into a portable device.The single powder identification or component analysis takes about 3 seconds.(5)Software for wheat phenotypic information analysis.Using Python and Flask toolkit,the B/S architecture-based user service software was developed,which adopted the way of data communication between browser and server to solve the compatibility problem with Windows,i OS,Android,Linux and other heterogeneous platforms.The designed program monitors the data request from the user,parses the parameters attached to the request,drives the sensing device to perform data collection,analysis and storage,and then combines the processing results to generate an HTML page for user.Users can control the developed leaf SPAD sensor,leaf water content sensor and powder analysis equipment through mobile browser,and access the collected spectrum/image data and corresponding results.