土壤磷素是植物最重要养分之一。 磷素在土壤中动态性强, 检测困难, 在可见-近红外光谱范围没有明显吸收波段, 因此研究基于其他光谱手段的磷素快速检测方法对于发展精细农业和智慧农业具有重要意义。 拉曼光谱具有受水分干扰小, 样本预处理小、 与红外光谱信息互补等特点, 国内外很多学者尝试了应用拉曼光谱对土壤磷素的检测。 但是, 拉曼信号弱, 稳定性差, 制约了拉曼光谱在土壤检测方面的应用。 为进一步弄清拉曼光谱与磷素的定量关系, 采用水溶性磷（KH2PO4）为研究对象, 研究了不同磷浓度的KH2PO4溶液对拉曼光谱产生的影响。 采用移动平均（MA）、 MA+基线校正（BL）、 MA+标准正态变量（SNV）、 MA+多元散射校正（MSC）对原始光谱（RS）进行预处理, 分析了低浓度（0.02~5 g·L-1）与高浓度（5.21~93.87 g·L-1）区间KH2PO4拉曼光谱的变异特性及其与磷浓度之间的关系, 建立了磷浓度含量的预测模型。 结果表明: （1）低浓度区间与高浓度区间光谱的变异系数具有显著差异, 高浓度区间光谱的离散程度较大; （2）低浓度区间的拉曼光谱未检测到明显的拉曼波峰, 浓度变化展现了明显的基线变化。 偏最小二乘（PLSR）模型决定系数R2=0.28~0.36; （3）高浓度区间的拉曼光谱在863与1 070 cm-1处检测到明显的拉曼波峰, PLSR建模结果为R2=0.65~0.7。 MA+SNV、 MA+MSC处理比MA单独处理模型预测精度高, 说明磷酸根的拉曼特征峰为模型主要贡献因子; （4）使用全浓度区间PLSR建模可增加PLSR模型精度（R2=0.73~0.89）。 使用RS建模的精度最高, 说明基线漂移对PLSR结果具有积极作用; （5）通过PLSR回归系数, 选取645、 863、 1 070和1 412 cm-1四点波段建立多元线性回归（MLR）模型, 决定系数R2接近1。 说明特征峰选取可以滤除背景光干扰, 抽取有效磷酸根浓度信号。 （6）由以上结果可知, 利用拉曼光谱定量检测水溶性磷的含量是可行的, 降低背景光干扰、 提高拉曼信号的稳定性的同时, 开发特征波段选择方法、 提高模型可重复性及抗干扰能力是高分辨率拉曼光谱检测技术的关键。

Soil phosphorus is one of the most important nutrients for plants. Phosphorus is highly dynamic in soil, and it is not easy to detect it. It has no obvious absorption band in the visible-near infrared range. Therefore, rapid phosphorus detection methods based on other spectral methods are of great significance for developing precision agriculture and/or smart agriculture. Raman spectroscopy has the characteristics of interference-free from water, less sample pretreatment, and complementary to infrared spectral information. Many researchers have tried to use Raman spectroscopy to detect soil phosphorus. However, the weak Raman signal and poor stability restrict its application in soil sensing. To clarify the quantitative relationship between Raman spectra and phosphorus, water-soluble phosphorus (KH2PO4) was used as a research target, and the effects of KH2PO4 solutions with different phosphorus concentrations on Raman spectra were studied. The raw spectra (RS) were processed by moving average (MA), MA+baseline correction (BL), MA+standard normal variable (SNV), and MA+multivariate scattering correction (MSC). Low concentrations (0.02~5 g·L-1) and high concentrations (5.21~93.87 g·L-1) of KH2PO4 and their relationships with Raman spectrum variation characteristics were analyzed. A prediction model for phosphorus concentration content was established. The results show that: (1) the coefficient of variation of the spectra in the low concentration range and the high concentration range were significantly different, and the dispersion degree of the spectra in the high concentration range was larger; (2) No obvious Raman peaks were detected in the Raman spectra of the low concentration range. Concentration changes exhibited significant baseline shifts. The coefficient of Determination (R2) of partial least squares regression (PLSR) models was 0.28~0.36; (3) Characteristic Raman peaks at 863 and 1 070 cm-1 were identified in the high concentration range, and PLSR modeling results were R2=0.65~0.7. The MA+SNV and MA+MSC treatments had higher prediction accuracy than the MA alone, indicating that the Raman characteristic peaks of phosphate radicals are the main contributing factors of the model; (4) PLSR modeling using the full concentration range can increase the prediction accuracy (R2=0.73~0.89). The modelling accuracy of using RS was the highest, indicating that the baseline shift has a positive effect on the PLSR results; (5) Through the PLSR regression coefficient, 645, 863, 1 070, 1 412 cm-1 were selected as characteristic bands to establish a multiple linear regression (MLR) model, and the coefficient of determination R2 was close to 1. It shows that the selection of characteristic bands can filter out the interference of background light and extract the effective features of phosphate variation. It can be seen from the above results that it is feasible to detect the content of water-soluble phosphorus by Raman spectroscopy quantitatively. Besides reducing the interference of background light and improving the stability of the Raman signal, a method for selecting characteristic bands is important to improve the repeatability and anti-interference ability of the model for high-resolution detection of Raman spectroscopy.