Soil nitrogen content forecasting based on real-time NIR spectroscopy

Abstract

Fast and precisely estimating total nitrogen (TN) content in soil helps to promote carrying out prescription fertilization. And soil moisture is a severe interference factor in forecasting soil nitrogen content based on real-time NIR spectroscopy. This paper aims at predicting soil nitrogen content based on real-time soil spectrum through exploring pretreatment method without artificial drying and sieving soil samples. Firstly, the real-time near infrared absorbance spectra of soil samples were measured and their characteristics were analyzed. Then 1st–7th level wavelet decompositions were carried out for each soil sample’s real-time spectrum. RSNR (Relative Signal-to-Noise Ratio) was constructed to evaluate wavelet filtering quality at different levels, and the results indicated that low-frequency signals obtained after the 3rd level wavelet decomposing had the best performance. And then 5 soil sample groups (each group had the same moisture content but different nitrogen contents) were selected and continuum-removal method was used for processing their filtering signals. And by using the methods combined wavelet analysis and continuum removal technology, six sensitive wavebands were determined for predicting the TN content in soil, which were 1375nm, 1520nm, 1861nm, 2100nm, 2286nm and 2387nm. Finally the real-time TN content detecting models were calibrated and validated based on PLSR (Partial Least Squares Regression) and SVM (Support Vectors Machine) respectively. For the PLSR model, its calibration R2 was 0.602 and its RMSEC was 0.051mg/Kg; the validation R2 was 0.634, the RMSEP was 0.056mg/Kg and its RPD=1.838. For the SVM model, its calibration R2 reached to 0.823, the RMSEC was 0.034mg/Kg, the validation R2 reached to 0.810, the RMSEP was 0.053mg/Kg and its RPD was 2.129. It showed that, by using the proposed approach in this paper, the interference of soil moisture was mostly removed from soil real-time spectrum in the process of soil total nitrogen prediction, and the TN content regression models established by using the six sensitive wavebands had great performances in predicting soil TN content in real time.