Estimation Of Soil Total-Nitrogen And Organic-Matter Using Near-Infrared Spectroscopy

The contents of soil total-nitrogen (TN) and soil organic-matter (SOM) are important parts of soil fertility. Monitor it timely which is the precondition of agricultural work going well. Traditional method of soil monitoring cannot meet the needs of production and management. However, the near-infrared (NIR) spectroscopy analysis technology has its characteristics such as rapid, real-time, non-destructive, which has been widely used in monitoring various fertility indexes of soil.In this paper, we compared various pretreatments and modeling methods. It can provide a certain theoretical basis of the efficient use of spectrum data to establish quantitative soil model. To be specific, we established the quantitative models of TN and OM content by using first derivative (FD), savitzky-golay (SG) convolution smoothing, multiplicative scatter correction (MSC), stepwise regression (SR), partial least squares regression (PLSR) and principal component regression (PCR). We collected109soil samples from the0-20cm of soil surface, which was near the borders of Jiangxia district and Wuhan city center. It was used for chemical analysis and spectroscopy analysis respectively. The NIR spectrometer was Antaris II Fourier transformer. First of all, there are three pretreatments named first-derivative, SG convolution smoothing, MSC to the original spectrum. After that, we respectively established PLSR models of TN and OM. The results showed that the effect of first derivative and SG convolution smoothing are better in the optimization of model. After spectral pretreatment of original spectrum (FD and SG convolution smoothing), we adopted two kinds of data processing combined spectral values and chemical values:(1)After original reflectance’s transformation (the first-order differential of reflectance’s logarithmic), we can find some characteristic bands which have the strongest correlation between TN/OM and NIR spectrum. After that, we used a modeling method named stepwise regression (SR) to establish the forecast equations using characteristic bands. Finally, we evaluated the accuracy of prediction equations;(2)We established the spectral prediction models of TN and OM using PCR or PLSR method, then tested the accuracy of models and compared them. Here are some conclusions in this paper: (1) Compared three different spectral pretreatments we found that the effects of FD and SG convolution smoothing are better in the optimization of model. It illustrated the necessity of the two pretreatments to improve the model’s accuracy.(2) In1000nm to2500nm range of NIR spectrum, after spectral pretreatment and transformation, the correlation between spectral reflectance and TN/OM content is enhanced. The best prediction spectral bands of TN/OM content are located at2262nm,2320nm in this paper. After first order differential of reflectivity’s logarithm, the optimal bands are used for establishing multiple stepwise regression models. We found that the effect of OM modeling is superior to TN in SR model.(3) There are good correlation between TN/OM content and NIR spectrum. The regression models can satisfy the requirements of accuracy. The determination coefficient (R2) are greater than0.5, and the relative prediction deviation (RPD) is greater than1.4. Moreover, the precision of OM’s PLSR model is the best (R2=0.84, the RPD=0.84). In general, the effect of PLSR’s modeling is better than PCR’s, and the effect of OM’s modeling is better than TN’s.(4) Whether it is establishing PCR’s model and PLSR’s model after a variety of spectral pretreatment, or established SR’s model after extracting the characteristic bands, the NIR spectrum analysis technology to do quantitative analysis of TN and OM content is feasible. There were different modeling effects and prediction accuracy in different modeling methods with the same kind of partition between calibration set and prediction set. In this article, the PCR/PLSR model is superior to the SR model.