不同土壤类型的有机质含量的可见-近红外光谱检测模型传递方法研究

利用可见-近红外光谱分析技术可以准确快速的获取土壤养分含量，但不同类型土壤间养分含量校正模型的普适性是亟待解决的关键问题。为提高有机质含量光谱校正模型在多类型土壤之间的普适性和农田在线检测有机质含量速度，利用美国M107B区66个样品建立基于可见-近红外光谱的土壤有机质含量的粒子群-最小二乘支持向量机（PSO-LSSVM）校正模型，预测M107B区的23个验证集样品的决定系数R2=0.859，相对分析误差RPD=2.660；将M107B区89个土壤样品作为校正集建模后对N116B区20个验证集样品的有机质含量预测，预测R2=0.562，预测RPD=0.952，模型的预测R2和预测RPD分别降低34.6%和64.2%，表明M107B区土壤有机质含量的可见-近红外光谱校正模型直接用于N116B区时，预测精度显著降低；将N116B区部分土壤样品加入到M107B区样品集后重新建模，并预测N116B区20个验证集样品的有机质含量，当加入的N116B区土壤样品数量达到35以上，预测R2>0.80，预测RPD>2.0；加入到校正集的N116B区土壤样品数量从0增加到50，模型预测R2从0.562增加到0.811，预测RPD从0.952增加到2.274，精度逐渐提高。结果表明，在M107B区校正模型中加入N116B区部分土壤样品建模，能够有效提高M107B区土壤校正模型对N116B区土壤有机质含量的预测精度；加入的N116B区土壤样品数量达到50以上，模型预测性能趋于稳定，预测精度达到实用要求，成功将M107B区土壤有机质含量校正模型传递给N116B区土壤；优先选择与M107B区土壤样品的有机质含量或光谱曲线差异较大的N116B区土壤样品参与建模，可有效避免模型传递时模型性能出现突变。提出的方法能够有效提高M107B区土壤的有机质校正模型对N116B区土壤的预测精度，为基于可见-近红外光谱的农田土壤有机质含量实时检测提供一种新的经济可行的模型传递方法，为提高多类型土壤的有机质含量检测模型的普适性提供一种有效的解决方案。

Research on Model Transfer Method of Organic Matter Content Estimation of Different Soils Using VNIR Spectroscopy

Soil properties can be estimated accurately and quickly using visible and near-infrared (VNIR) diffuse reflectance spectroscopy. However, a key problem is the lack of universal nutrient content calibration models for different soils. To improve the universality of the soil organic matter (SOM) content calibration model for different types of soils and the speed of online detection of the SOM in farmland, sixty-six samples of soil from M107B in the United States were used to establish the SOM content. Calibration model using the particle swarm optimization-based least squares support vector machines (PSO-LSSVM) method using VNIR spectroscopy. Then this calibration model predicted 23 samples of the validation set from M107B. The results gave the coefficient of determination (R2) and the ratio of standard deviation to root mean square error of prediction (RPD) of 0.859 and 2.660, respectively. Subsequently, we predicted the SOM content of the validation set, including 20 samples from N116B, by the PSO-LSSVM calibration model of all 89 soil samples from M107B. The results showed decreases in the R2-value (0.562) and RPD (0.952). These decreases in R2 and RPD values by 34.6% and 64.2%, respectively, indicated that the prediction accuracy was significantly decreased when the PSO-LSSVM calibration model of SOM content in M107B was directly used to predict SOM content in N116B. The PSO-LSSVM calibration model established by the calibration set, a combination of some soil samples from N116B and all 89 samples from M107B was also used to predict SOM content of the previous validation set from N116B and gave the R2 values that were more than 0.80 and RPD values that were more than 2.0 when the number of soil samples from N116B was added over 35. In addition, R2 increased from 0.562 to 0.811. RPD increased from 0.952 to 2.274 when the number of soil samples from N116B added to the calibration set increased from 0 to 50. The results showed that calibration model accuracy could be effectively improved by adding some soil samples from N116B to M107B calibration set when predicting SOM content in N116B. The prediction performance of models was stable, whereas the prediction accuracy met practical requirements when the number of soil samples from N116B added to the calibration set was more than 50. In addition, the calibration model of SOM in M107B was successfully transferred to the soil in N116B, and the samples in N116B with large differences in organic matter content or spectral curve from samples in M107B are preferred to adding to the calibration set because this method can effectively avoid the mutation of model transfer performance. In conclusion, the results provided a method to improve the SOM prediction accuracy of N116B soil using the SOM calibration model of M107B soil. Furthermore, the results provided a new, economical and feasible model transfer method for real-time estimating of SOM content in farmland based on VNIR. The results also provided an effective solution to improve the universality of the SOM content calibration model for different soil types.