Research On Estimation Algorithm Of Leaf Water Content Of Tomato Under Water Stress Based On Hyperspectral Imaging

Water stress is one of the key abiotic stresses that affect plant growth and crop yield,accurate and rapid understanding of crop water status is important for crop health assessment and farmland irrigation management.Crop water conditions can affect its physiological and biochemical indicators,and the overall water status of the crop can be evaluated by monitoring the water content of the crop leaf.The hundreds of narrow spectral bands of hyperspectral can sense subtle changes in crops,providing a new technical mean for dynamic non-destructive testing of leaf water content.This study focuses on potted tomatoes and investigates methods for identifying water stress and estimating leaf water content based on hyperspectral imaging technology,providing technical support for rapid and non-destructive monitoring of crop water status.The main research contents are as follows:(1)Use the hyperspectral imaging system to collect images of tomato leaves processed in different experiments;Eliminate background information through image processing technology,locate the area where the leaves are located,and export black and white corrected hyperspectral data from ENVI;Use SG smoothing to eliminate noise;Use analytical balance and drying oven to measure the water content of leaves.(2)Tomato water stress classification model based on band matrix reduction method.Using clustering algorithms and statistical histograms to obtain the spectral characteristics of tomato leaves under different water conditions.The discrimination coefficient Coef(i)is calculated from the average spectrum and the first-order derivative spectrum to determine the importance of bands,and a band matrix is constructed based on band sorting.Combining the idea of recursive feature elimination and convolution in neural networks,the band matrix reduction method is proposed to reduce the band matrix elements.Finally,cross validation(CV)is used to compare the classification accuracy of all band combinations,the band combination with the maximum value is the optimal band subset.The experimental results show that compared with commonly used band selection methods,this research method not only improves the recognition accuracy of water stress,but also has a significant dimensionality reduction effect.(3)Research on the estimation model of leaf water content based on DD network.First,the correlation between leaf water content and spectral reflectance is analyzed,and hierarchical clustering is used to eliminate multicollinearity in hyperspectral data,the idea is to divide the bands with high correlation into the same cluster,and the correlation between different clusters is lower.Draw specific lines(y=1-correlation)in the tree graph showing the hierarchical clustering process to obtain the band clustering results and retain the band most related to water content in each cluster.Then,the DD network is used to construct the leaf water content estimation model.Finally,the constructed dataset was used for leaf water content estimation experiments and compared and analyzed with commonly used methods such as partial least squares regression(PLSR)and BP neural network.The results showed that the correlation between the estimated value of DD network and leaf water content was the highest(0.845)after eliminating multicollinearity.(4)Research on automatic construction method of vegetation index(VI)based on relationship spectrum to estimate leaf water content.After eliminating multicollinearity,add the difference and sum between any two bands as the input of DD network.In addition,add a point multiplication module in front of the whole network.The network can comprehensively consider the problems to be solved in VI automatic construction,including band selection,wavelength weight and equation form.Substitute the weight matrix obtained from DD network training into MATLAB to obtain the network expression and generate a relational spectrum,the spectral vegetation index is composed of dividing the term with the highest absolute value of the coefficient and the term with the lowest absolute value in the expression.The experimental results show that the correlation between spectral vegetation index and leaf water content is 0.77,which is much higher than PLSR and common vegetation indices related to leaf water content,and better than the normalized difference spectral index(NDSI)and ratio spectral index(RSI)obtained through exhaustive method,providing a new approach for VI construction and rapid detection of crop water status.