土壤养分直接关系到作物产量与品质状况, 然而传统化学方法检测存在化学试剂消耗大、 耗时费力等问题, 不能满足精细农业的需求。 快速获取土壤养分信息是发展精细农业、 绿色农业的关键, 想要了解土壤肥力状况, 必须先了解有机质和总氮的含量状况。 许多研究表明, 长波近红外光谱被广泛应用于土壤检测领域, 然而短波可见/近红外光谱在土壤有机质和总氮的研究上却非常罕见。 以江西省吉安市安福县和南昌市新建区的四个村庄作为研究区, 根据2×2网格法采集了深度为10～30 cm的棕壤、 红壤和水稻土三种最为典型的土壤样品共180份。 经过研磨、 风干等处理后用四分法均匀划分为两份, 用于测定样品光谱信息和理化信息。 将土壤样品按照2∶1(120∶60)划分为建模集和预测集。 考虑到首尾端波段噪声较大, 故去除325～349和1 051～1 075 nm波段, 将350～1 050 nm波段用于光谱分析。 通过连续投影算法(SPA)筛选出有机质12个特征波长点, 总氮11个特征波长点, 考虑到土壤光谱信息与土壤理化性质之间可能存在非线性联系, 建立全波段与特征波长的线性偏最小二乘回归(PLSR)模型和非线性最小二乘支持向量机(LS-SVM)模型对土壤有机质和总氮进行研究, LS-SVM模型采用两步网格搜索法优化了两个超参数γ和σ2。 研究结果表明: (1)土壤的光谱反射率随波长增加反射率升高, 反射率曲线中460、 550、 580、 740和900 nm处有较为明显的吸收特征。 (2)从PLSR模型和LS-SVM模型结果分析可知, 非线性模型LS-SVM具有更好的预测精度, 分析认为土壤光谱信息与土壤理化性质之间存在一些非线性关系。 (3)通过连续投影算法筛选的特征波长提高了模型精度, 优化了模型运行效率。 SPA-LS-SVM模型是所有模型中最优的预测模型, 其中有机质模型的R2pre为0.884 7, RMSEp为0.104 8, RPD为2.945 0, 总氮模型的R2pre为0.901 8, RMSEp为0.010 4, RPD为3.191 1。 (4)本研究说明可见/近红外光谱能够用于测量不同类型的土壤有机质和总氮含量, 并且达到较好的预测效果。 可见/近红外光谱在土壤检测领域具有巨大潜力。

Soil nutrient status is directly related to crop yield and quality. However, traditional chemical methods have problems such as large consumption of chemical reagents, being time-consuming and labor-intensive, and cannot meet the needs of precision agriculture. Quickly obtaining soil nutrient information is the key to developing precision and green agriculture. To understand soil fertility, one must first understand the content of organic matter and total nitrogen. Many studies have shown that near-infrared spectroscopy is widely used in soil detection, but visible/near-infrared spectroscopy is very rare in the study of soil organic matter and total nitrogen. Taking four villages in Anfu County, Ji’an City, Jiangxi Province, and Xinjian District, Nanchang City as the study areas, the three most typical soil samples, brown soil, red soil and paddy soil, with a depth of 10~30 cm were collected according to the 2×2 grid method180 share. After grinding, air-drying, etc., the samples were divided into two parts by the method of quartering, which was used to determine the samples’ spectral and physicochemical information. The soil samples were divided into modeling set and a prediction set according to 2∶1 (120∶60). Considering the large noise in the first-end band, the 325~349 nm and 1 051~1 075 nm bands were removed the remaining 350~1 050 nm band was used for spectral analysis. 12 wavelength points of OM and 11 wavelength points of TN were screened out by successive projections algorithm. Considering the possible nonlinear relationship between soil spectral information and soil physical and chemical properties, a full-band, the linear partial least squares regression (PLSR) model of characteristic wavelengths and the nonlinear least squares support vector machine (LS-SVM) model were used to study soil organic matter and total nitrogen. The LS-SVM model was optimized by a two-step grid search method. Two hyperparametersγ and σ2. The results show that: (1) The spectral reflectance of soil increases with the increase of wavelength, and the reflectance curve has obvious absorption characteristics at 460, 550, 580, 740 and 900 nm. (2) From the analysis of the results of the PLSR model and the LS-SVM model, it can be seen that the nonlinear model LS-SVM has better prediction accuracy, which may be due to the nonlinear relationship between soil spectral information and soil physical and chemical properties. (3) The characteristic wavelength screened by the continuous projection algorithm improves the model accuracy and optimizes the model operation efficiency. The SPA-LS-SVM model was the best predictive model among all the models, among which the R2pre of the organic matter model was 0.884 7, the RMSEp was 0.104 8, and the RPD was 2.945 0. The R2pre of the total nitrogen model was 0.901 8, the RMSEp was 0.010 4, and the RPD was 3.191 1. (4) This study shows that visible/near-infrared spectroscopy can measure different types of soil organic matter and total nitrogen content, achieving better prediction results. Visible/NIR spectroscopy has great potential in the field of soil detection.