| Soil total nitrogen (TN), organic matter (OM), available potassium (K) and available phosphorus (P) are the main nutrients for crop growth and soil measurement important parameters. Conventional measurement methods of these parameters require expensive testing equipment and have many disadvantages such as high cost problem and low efficiency, which hold back the development of precision agriculture management. And the modern agriculture requests that the soil nutrition measurement should be effective. Near infrared spectroscopy analysis as a reliable, rapid, little sample preparation requirement, low-cost, convenient, nondestructive and green technique becomes more and more important in the area of soil nutrition measurement. Near infrared spectroscopy are highly sensitive to C-H, O-H and N-H bonds of soil components such as total nitrogen (TN), organic matter (OM) making their use in the agricultural and environmental sciences particularly appropriate. The analytical abilities of near infrared spectroscopy depend on the repetitive and broad absorption of light by C-H, O-H and N-H bonds.The main creative results were achieved as follows:(1) Visible near infrared spectroscopy combined with genetic algorithm and successive projections algorithm was investigated for soil organic matter (OM). In order to simply calibration model, a total of18characteristic wavelengths were selected by using genetic algorithm and successive projections algorithm and18characteristic wavelengths were subjected to partial least squares regression (PLSR) with leave-one-out cross validation to establish calibration models of soil organic matter (OM) with coefficient of determination (R2) of0.81,0.83, RMSEP of0.22,0.20and residual prediction deviation (RPD) of2.31,2.45for the calibration set and prediction set respectively. The results showed that using genetic algorithm and successive projections algorithm can simply the model greatly.(2) Hyperspectral imaging technology is a rapid, non-destructive, and non-contact technique which integrates spectroscopy and digital imaging to simultaneously obtain spectral and spatial information. Hyperspectral images are made up of hundreds of contiguous wavebands for each spatial position of a sample studied and each pixel in an image contains the spectrum for that specific position. With hyperspectral imaging, a spectrum for each pixel can be obtained and a gray scale image for each narrow band can be acquired, enabling this system to reflect componential and constructional characteristics of an object and their spatial distributions. Hyperspectral imaging technology combined with least-squares support vector machine (LS-SVM) was investigated for soil organic matter (OM). A near-infrared hyperspectral imaging system (874-1734nm with256bands) was established to acquire the hyperspectral images of the samples. A region of interest (ROI) pixels of the hyperspectral image of each sample was defined, and the average reflectance spectrum of the ROI was extracted. To remove the absolute noises of the spectra, only the spectral range951-1713nm was used for analysis, and the extracted180reflectance spectra were preprocessed by Savitzky-Golay smoothing (SG), MSC, and Wavelet Transform (WT) methods. The preprocessed spectra were then used to select sensitive wavelengths by Successive Projections Algorithm (SPA) and Genetic Algorithm-partial least squares (GA-PLS) methods. Different numbers of sensitive wavelengths were selected by different variable selection methods with different preprocessing methods. Partial least squares (PLS) was used to build models with the full spectra,and back-propagation neural network (BPNN) and LS-SVM were applied to build models with the selected wavelength variables.The overall results showed that BPNN and LS-SVM models performed better than PLS models, and the LS-SVM models with the selected wavelengths based on MSC preprocessed spectra obtained the best results with the determination coefficient (R2),RMSEP and RPD were0.82,0.27and2.37for calibration set, and0.78,0.29and2.24for the prediction set. The MSC preprocessing method showed the best performance in all PLS, BPNN, and LS-SVM models. The results indicated that it was feasible to use near-infrared hyperspectral imaging to predict soil organic matter content.(3) This study proposed a new method using visible and near infrared (Vis/NIR) hyperspectral imaging for the detection of soil type. In this study, a hyperspectral imaging system (380-1023nm) was developed to perform classification of soil type based on gray level co-occurrence matrix (GLCM) and least squares support vector machines (LS-SVM). Altogether150soil samples were collected for hyperspectral image scanning, and mean spectra were extracted from the region of interest (ROI) inside each image. LS-SVM was applied as calibration method to correlate the spectral and GLCM data for the100samples in calibration set. Then the LS-SVM model was used to predict50prediction samples. Spectra of soil sample was extracted from region of interest (ROI) and competitive adaptive reweighted sampling (CARS) algorithm was used to select the key variables from near-infrared hyperspectral imaging data and principal component analysis (PCA) was performed with the goal of selecting the first principal component (PC) image that could potentially be used for classification system. Then,12texture features (i.e., mean, standard deviation, smoothness, third moment, uniformity, and entropy) based on the statistical moment were extracted from PC1image. Finally,12gray level co-occurrence matrix (GLCM) variables combined with31characteristic wavelengths for each soil sample were extracted as the input of LS-SVM. Experimental results showed that discriminating rate was100%in the prediction set. The results indicated that hyperspectral imaging technology combined with chemometrics and image processing allows the classification of soil type.(4) Visible near infrared spectra technology was adopted to detect soil total nitrogen content. Raw spectra and wavelength-reduced spectra with six different pretreatment methods were compared to determine the optimal wavelength range and pretreatment method for analysis. Spectral variable selection is an important strategy in spectrum modeling analysis, because it tends to parsimonious data representation and can lead to multivariate models with better performance. In order to simply calibration models, the wavelength variables selected by three different variable selection methods (i.e. regression coefficient method (RC), successive projections algorithm (SPA) and genetic algorithms-partial least squares analysis (GA)were proposed to be the inputs of calibration methods of PLS, MLR and LS-SVM models separately. These calibration models were also compared to select the best model to predict soil TN. Both linear calibration algorithms of MLR and PLS and non-linear calibration algorithm of LS-SVM obtained similar results based on three variable selection algorithms. The best results indicated that PLS, MLR and LS-SVM obtained the highest precision with determination coefficient of prediction r2pre=0.81, RMSEP=0.0031and RPD=2.26based on wavelength variables selected by RC and SPA as inputs of models. The overall consequence showed that Vis-NIR technology can be used to measure the content of soil total nitrogen, RC and SPA as two variable selection methods were very useful in spectra analysis, and it could perform well with less input dimension and computation complexity in the soil total nitrogen estimation.Visible near infrared spectroscopy combined with uninformative variable elimination (UVE) and wavelet (WT) algorithm was investigated for soil total N (TN). In order to simply calibration model, a total of46characteristic wavelengths were selected by using UVE and WT algorithm and these46variables were used as input to partial least squares regression (PLSR) with leave-one-out cross validation to establish calibration models of soil total N (TN) with coefficient of determination (R2) of0.87and residual prediction deviation (RPD) of3.08for the prediction set. The results showed that using UVE and WT algorithm can simply the model greatly.(5) Portable short wave NIR spectroscopy technology was used to measure soil total nitrogen. The reflectance spectra were preprocessed by Savitzky-Golay smoothing (SG), Reduce (RD), and Wavelet Transform (WT) methods. The preprocessed spectra were then used to select sensitive wavelengths by competitive adaptive reweighted sampling (CARS), Random frog and Successive Projections Algorithm (SPA) methods. Different numbers of sensitive wavelengths were selected by different variable selection methods with different preprocessing methods. Partial least squares (PLS) was used to build models with the full spectra, and ELM and LS-SVM were applied to build models with the selected wavelength variables. The overall results showed that PLS and LS-SVM models performed better than ELM models, and the LS-SVM models with the selected wavelengths based on SPA obtained the best results with the determination coefficient (R2), RMSEP and RPD were0.63,0.0079and1.58for prediction set. The results indicated that it was feasible to use portable short wave near-infrared spectral technology to predict soil total nitrogen.(6) Visible near infrared spectroscopy combined with Monte Carlo Uninformative Variables Elimination (MC-UVE) and genetic algorithm (GA) was investigated for soil available K (K). In order to simply calibration model, a total of50characteristic wavelengths were selected by using MC-UVE and GA and these50variables were used as input to partial least squares regression (PLSR) with leave-one-out cross validation to establish calibration models of soil available K (K) with coefficient of determination (R2) of0.68, RMSEP of6.45and residual prediction deviation (RPD) of1.7for the prediction set. Visible near infrared spectroscopy was investigated for soil available P by using competitive adaptive reweighted sampling (CARS) algorithm. In order to simply calibration model, a total of26characteristic wavelengths were selected by using CARS algorithm and these26variables were used as input to partial least squares regression (PLSR) with leave-one-out cross validation to establish calibration models of soil available P with coefficient of determination (R2) of0.64, RMSEP of3.8and residual prediction deviation (RPD) of 1.67for the prediction set. The variables selected by CARS algorithm can be used to develop model to predict soil available P.(7) The soil total nitrogen measurement portable instrument was based on USB4000and near infrared spectroscopy technology. The instrument consisted of two parts, software section and hardware section. The hardware part included instrument box, portable power, lamp source, Y type optical fiber, driven circuit, main board, touch screen and AID circuit. The software part includes model load module, spectra collection module, spectra save module, result display module and parameters set module and so on. Portable instrument was developed using JAVA language and develop kit was supported by USB4000manufacturers. When detecting soil total nitrogen, firstly, measurement soft was set to start automatically. Secondly, open lamp source and collect fiber, set the corresponding testing parameters and testing conditions, click on the touch screen to begin collecting buttons. The reflectance spectral curve and the detection results was acquired and displayed in the touch screen.The above results realized the fast and high precision detection of soil OM, N, P and K. They also supplied theoretical basis of detection instruments for the determination of soil nutritional information which had a promising application prospect. |
PDF Full Text Request