Research On Plant Feature Information Extraction Using Spectroscopy And Hyperspectral Imaging Technique

Stress conditions such as moisture-deficient, cholorothyll-deficient and diseases, have a high impact in the growth and production of the plant. A fast, efficient and non-destructive detection of plant physiology such as aging, damage, environmental stresses and other mineral nutrition deficiency is the precondition of implementing precision management in agriculture. Recently, spectroscopy has been becoming one of the most important modern analysis techniques in angricultural producing at the advantanges of fast, nondestructive, pollution-free and low-cost. In this thesis, based on spectral analysis and hyperspectral imaging technique, we discussed the instruments and methods of detecting the nutritional contents of winter wheat, the methods of the citrus greening disease detection and the algorithms of blueberry fruits identification and yield estimation. It aimed to provide a theoretical basis for farmland management, pest control and crop yield estimation.The major works included:(1) Development of portable plant components measurement instrument based on near-infrared spectroscopy. A portable and multi-function plant components measurement instrument was developed, which was mainly used to measure the contents of moisture and cholorophyll of winter wheat leaves. Based on the method of Dual-wavelength, the integral type leaves clamp was designed, which was composed of double LED, spectra filters, lens, and photoelectrical receiving device and so on. The design simplified the optical structure, enhanced the stability of apparatus, and further improved the signal-to-noise ratio and data precision. The hardware mainly contained four parts, including micro-control system, driving circuit of light source, photoelectric conversion and signal conditioning circuit, and LCD display circuit. In addition, an open data interface was reserved in the instrument for an easy connection to the leaves clamp, which can achieve a real-time, quick and nondestructive measurement of different plant components (680nm and940nm used for chlorophyll measurement,880nm and970nm used for moisture measurement). By the calibration experiments, the equipment is stable and easy to use.(2) Detection of chlorophyll and moisture contents of winter wheat leaves using spectral analysis. Moisture and chlorophyll contents in the different growth stages of winter wheat leaf were studied based on vis-nir spectroscopy. The different spectral preprocessing methods such as normalization, first derivative (FD), move smoothing, Savitzky-Golay smoothing (SG), multiplicative scatter correction (MSC) and variable standardization correction (SNV) were compared in the quantitative analysis model. And the performances of the models building with multiple linear regression (MLR), partial least squares regression (PLSR) and artificial neural network (ANN) were evaluated based on the full spectra, principal components (PCs) and optimal characteristic bands. The results indicated that the model using various spectral pretreatment had a higher prediction accuracy, and based on principal components and the optimal spectral bands MLRand PLS had a better prediction ability.(3) Identification of citrus greening using spectral analysis and hyperspectral imaging. First, the spectral feature extraction was studied based on vis-nir spectroscopy. The values of discriminability were used in the extraction of spectral features, which reduced the standard deviation of classification model. By the Fisher linear discriminant analysis and classification tree, the model based on the spectral features yielded more than88%accuracies. Compared the results of classifers k-nearest neighbor (KNN) and Naive Bayes, that spectral features were treated as input variables was significantly superior to the original spectrum, which confirmed the importance of spectral feature selection. And then spectral feaure and textural feature were used for HLB identification based on hyperspectral imaging technique. Based on the gray histogram and gray-level co-occurrence matrix (GLCM), the textural features can significantly improve the accuracies, indicating that the hyperspectral imaging containing both spectral information and spatial textural information had a great potential for citrus greening detection.(4) Identification of blueberry fruits of different maturity using hyperspectral images. Small-size fruits and complicated background were always a big problem for blueberry fruits detection. This paper collected hyperspectral images of different varieties of blueberry under different weather conditions in field, preprocessed them using the methods of atmospheric correction, background subtraction, multivariate regression and so on to normalize the images. Using the joint algorithm of support vector description (SVDD) and k-menas clustering, the fruits with different maturity in the same branch were classified. Results were compared with the classical detection methods KNN and spectral angle mapper (SAM), SVDD was adapt to complex background and had higher detection accuracies, which achieved89.5%,85.9%and84.1%respectively for the three different growth stages (mature, intermediate, young) blueberry fruits. The study showed the joint algorithm of SVDD and k-menas clustering had a good potential in blueberry fruit identification, which provided a good basis of grove management and yield estimation for fruit growers.