总氮是衡量土壤肥力的重要成分指标。 传统的检测土壤总氮含量的化学方法操作复杂且费时费力, 采用傅里叶近红外（FT-NIR）对土壤总氮的含量实现直接快速定量分析； 然而, 利用偏最小二乘（PLS）等线性分析方法定量预测土壤样本的总氮含量, 定标预测模型有可能被理想化, 不利于在线检测的实际应用。 考虑给定量分析模型添加容错机制, 将FT-NIR定量分析转化为半定性判别分析, 以加强光谱模型的应用能力, 提出区间间隔搜索主成分分析逻辑回归（iPCA-LR）方法, 结合PLS的先验定量预测值, 通过设定r=0.05, 0.10, 0.15三个不同的容错阈值范围, 给样本赋予先验判别标记, 将定量分析模式转换为半定性判别模式, 建立土壤总氮的FT-NIR半定性判别模型, 同时, 对比讨论基于k＝5, 10, 15, 20四种不同子波段数量的区间划分数据的潜变量转换模式, 优选FT-NIR光谱特征子波段, 并讨论优选连续子波段的组合建模情况。 结果表明, 不同阈值范围下的FT-NIR半定性判别模型的预测准确率差别较大, 但不同阈值范围的最优判别模型的预测准确率均在75%以上, 各个区间划分的优选子波段或合并子波段的判别准确率均达到了90%以上, 可以满足不同程度的应用水平。 利用PLS结合iPCA-LR将定量预测转换为半定性判别的方法能够应用于土壤总氮的FT-NIR光谱分析, 能够解决常规PLS定量分析容易过拟合和过于理想化的问题, 半定性判别结果更符合实际, 有利于光谱技术的在线应用。

Nitrogen is an important component to measure soil fertility. The traditional chemical method for detecting soil nitrogen content is complex and time-consuming. Fourier transform near infrared (FT-NIR) technology is utilized for direct and rapid quantitative determinationof soil nitrogen. Nevertheless, the calibration models always perform too ideally well to believe when established by the linear analytical methods, like partial least squares (PLS). That is not convinced for the practical application in on-line detection. In this paper, we proposed a fault-tolerant mechanism to be plug-into the quantitative analytical model, transforming the FT-NIR quantitative mode into a quasi-qualitative discriminant mode. In this way, the application ability of the calibration model can be enhanced. A new discriminant method was proposed for quasi-qualitative determination by combining the interval search principal component analysis algorithm with logistic regression (iPCA-LR). The nitrogen contents of soil samples were firstly predicted based on the common PLS regression. The fault-tolerant threshold was set as three different values of 0.05, 0.10 and 0.15, respectively. The samples were marked as accurately or non-accurately discriminated according to the priori predictive values and the thresholds, so that the original quantitative calibration method was transformed into a new quasi-qualitative discriminant method. The iPCA-LR method was applied for the FT-NIR quasi-qualitative discrimination of soil nitrogen. In the same process, we also discussed the latent variable extraction based on different wavebands that were generated by tuning the waveband division number as 5, 10, 15 and 20. Some informative FT-NIR wavebands were selected with optimal discriminant accuracy. And some combination of informative wavebands were also tested. Results showed that the FT-NIR quasi-qualitative discriminant predictive accuracy varied significantly for different thresholds, but fortunately the worst optimal accuracy climbed tothe level slightly above 75%. And the test of different informative wavebands or the combination of informative wavebands output optimal calibration models with the accuracy above 90%. These results were able to meet some practical cases of online detection. In the application of FT-NIR prediction of nitrogen content in soil samples, the proposed method of iPCA-LR manage to transform the common quantitative prediction problem into the quasi-qualitative discriminant problem when combined with the priori PLS prediction. The newly proposed method deals with the disadvantages of overfitting and overidealistic modeling that always appears in common PLS quantitative analysis. In comparison, the quasi-qualitative discriminant mode is more suitable for actual cases in field detection, more beneficial for real-time application of spectroscopy technology.