Nitrogen Detection Of Lettuce With Multi-source Imaging Based On THz-NIR Hyperspectral And 3D Light Field

Facility agriculture is developing rapidly,in which facility vegetables are one of the main sources of vegetable supply.Scientifically cultivating crops and developing efficient agriculture are conducive to promoting the modernization of agricultural production in China.To ensure the yield and supply of vegetable crops,rational and precise fertilization is needed in the process of crop growth.The prerequisite for achieving this goal is to be able to quickly and accurately obtain the nutritional status of crop growth.However,traditional methods relying on expert experience and chemical determination are time-consuming and laborious,and are not suitable for the development of efficient agriculture.In this paper,taking lettuce as the research object,aiming at the urgent requirement of high-precision nondestructive detection of crop nutrition status,terahertz,hyperspectral and light field imaging technology were applied,combined with the extraction and optimization of internal and external,2D,3D,multi-source and multi-scale imaging features,and the terahertz and hyperspectral prediction models and their visual feature images of lettuce nitrogen nutrition detection were established respectively.Based on the light field imaging data,the lettuce nitrogen nutrition growth model was established,and a mobile platform was built.It provides a reference for nondestructive detection of crop nutrition and is of great significance for precise fertilization.The main research contents and conclusions of this paper are as follows:(1)Using soilless culture,four gradient levels were set to cultivate lettuce samples under nitrogen stress.The relevant data of lettuce samples were collected by terahertz,hyperspectral and light field imaging systems,and the true value of nitrogen content of lettuce samples was measured by chemical methods.(2)Analyzing the terahertz data of samples,the research range was selected to be0.2?1.2 THz.After preprocessing the data by SG smoothing and MSC algorithm,SPXY algorithm was selected to divide the sample set,and SCARS,i PLS and IRIV algorithm were used to extract the feature frequency band of data,and LS-SVM and KELM terahertz prediction models of lettuce nitrogen based on feature frequency band were established.Among them,the LS-SVM prediction model with SCARS feature in RBF kernel of power spectrum dimension was best,the determination coefficient and root mean square error of the prediction set are 0.9606 and 0.1986 respectively;SCARS feature of power spectrum dimension is also the best in KELM prediction model,the determination coefficient and root mean square error of prediction set are 0.9596 and0.1997,respectively.Through the established optimal model,the visual expression of terahertz sample features is realized.(3)Analyzing the hyperspectral data of samples,the research range was selected to be 1000?1600 nm.After SG smoothing,SNV algorithm pretreatment and SPXY algorithm sample set division,the feature wavelength of data is selected by RF,SPA and ICO algorithm,and the hyperspectral prediction model of lettuce nitrogen based on LS-SVM and KELM is established.Among them,the ICO feature of RBF kernel is the best in LS-SVM prediction model,the determination coefficient and root mean square error of prediction set are 0.9603 and 0.2620,respectively;ICO feature is also the best in KELM prediction model,the determination coefficient and root mean square error of prediction set are 0.9620 and 0.2628,respectively.Through the best model,the visual expression of hyperspectral sample features is realized.In order to obtain a more comprehensive and effective detection method,combining the advantages of terahertz and hyperspectral features,a fusion feature model is established by Res Net-18 architecture with few-shot learning.The results show that the integration of terahertz SCARS features with hyperspectral ICO features is the best,and the accuracy of correction set and prediction set is 90.25% and 85.24%,respectively.(4)The noise reduction and simplification methods of point cloud data are introduced.The growth parameters of samples are obtained by calculation,and the fitting model of growth parameters and the prediction model of nitrogen growth characteristics of lettuce based on multiple linear regression and nitrogen content are established respectively.The determination coefficients of correction set and prediction set are 0.9391 and 0.9474,and the root mean square error is 1.0298 and 0.6557,respectively.The research shows that the growth features model established by point cloud data has certain prediction effect.A mobile platform which can carry a variety of nutrition and growth detection equipment was built,and the main functions and components of the platform were introduced.