| The content of soil organic matter, available N and P is a major limiting factor for the growth of crops. To implement precision fertilization, soil nutrients need to rapid detection.Because lack of fast determination method and actual instrument, Lou soil was prepared for research by portable spectrometer. After analyzed the relationship of spectrum and soil components, we compared different pretreatment methods and established the optimization mathematics model for measurement the organic matter, the available p and n content,provided the basis for real-time airborne determination. Main contents and conclusions are as follows:(1)Proved the optimization height. At first, 160 soil samples were collected both in Northwest A & F University and around corn field, then obtained the soil sample without grinding spectrum data range of 900-1700 nm in different heights(5, 7, 10, 12 and 15cm)using portable near infrared spectrometer, established soil organic matter content forecast model with PLS method. The results showed that forecast model works best when the sampling height was 10 cm.(2)For spectral and content data respectively, using PCA score and 3 times the standard deviation criteria to determined and eliminated anomalous one, the results showed that after removal the prediction accuracy has improved significantly. Then using Random Division,Kennard-Stone Division and sample set partitioning based on joint x-y distance methods divided the spectrum excluding the spectral data classification of abnormal samples, the results of prediction model of soil organic matter content showed that the SPXY sample method was better.(3) Because of the spectral data wavelength redundancy and spectrum peaks overlapping, it need to find effective wavelength selection.The effect of four different wavelengths selecting methods(SPA, CARS, sCARS and Random Frog) were analyzed and proved the optimum model. Then established the PLS model with sensitive wavelength to verify the better one, and the results of soil organic content present when by the SPA algorithm the Rc2 was 0.8559 slightly decreased by 4.9% compared with the whole-band wavelength of 0.8982 while the Rp2 was 0.7952 better than whole-band model; For the prediction model of the available p content indicated that the sCARS method can filter out96.25% of the redundant information and the modeling result is the best satisfied.(4)Chosen MLR PCR and PLS in the linear analysis methods and nonlinear analysis methods of RBF neural network, WNN and LSSVM models to comparative analysis, and the results shown that the coefficient of PLS prediction model is 0.8019, root mean square error is0.1794; the prediction coefficient Rp2 and the RMSEP of RBF neural network modeling are0.8281 and 0.1646 respectively, two kinds of model both have high precision, and can accomplish quick prediction. For the available P content measurement, the LSSVM model has better effect with its coefficient Rp2 was 0.8325 much better than RBF 0.8130 by 2.4%increased, and WNN 0.7450 by 10.5%, the RMSEP of LSSVM was 8.4733. For the available N content measurement, using method of linear and non-linear models, the comparison found that predictive validity cannot meet the actual application requirements. The model experiment shown that the linear model is simple construction and better popularization, but the accuracy is not satisfied and poor robustness, while the sample range of the RBF neural network is expansive and more adapted to practical application.(5)In order to improve the prediction accuracy, it was established the DBN model to predict soil available p content.After the modeling it found that the Rp2 of DBN prediction model is 0.8879, obviously better than the WNN results by 16.1%, and the RMSEP dropped only 1.7166, which fully demonstrates the DBN has the characteristics of better initial values and the automatic learning greatly simplifies the model, shorter the computing speed and better to solve the problem of the modeling dissemination. |
PDF Full Text Request