基于样品复配的奶牛场粪水近红外光谱氮磷速测模型构建

为提高样品代表性,构建近红外光谱的理想校正模型,实现对奶牛场粪水氮磷含量的速测,本研究基于粪水运移过程代表性暴露位点的原始样品,按照不同比例进行样品复配,填补现场不易采集到的"黑箱"位点样品,运用偏最小二乘法构建了基于最优光谱预处理方法的原样模型、复配模型和融合模型。结果表明:相比原始样品,原样+复配样品总氮、总磷的变异系数分别降低了0.103、0.107,提升了浓度分布均匀性,丰富了粪水光谱信息。相比原样模型,融合模型总氮和总磷的决定系数(R²_pred)分别提升了0.049和0.061,相对分析误差(RPD)分别提升了1.547和0.176。相比复配模型,融合模型总氮和总磷的R²_pred分别提升了0.026和0.022,RPD分别提升了0.470和0.052。验证结果表明,总氮和总磷模型的R²_pred分别为0.903和0.878、RPD分别为2.916和2.508。研究表明,样品复配的方法可有效提高校正集样品的代表性,提升模型预测性能,为还田前粪水养分的快速定量提供技术支撑。

Rapid determination modeling of slurry nitrogen and phosphorus from dairy farm based on samples compounding

In order to improve the representativeness of samples, an ideal correction model of near infrared spectroscopy (NIRS) was constructed to realize the rapid measurement of nitrogen and phosphorus content in slurry of dairy farm. The study was counted on the original samples that gathering from representative uncovered points throughout slurry movement process in dairy farm. Compound samples were made in terms of different proportions to fill in the "black box" samples which were not easy to be collected. Applying the partial least square, original models, compound models and fusion models were respectively built up relying on the optimal spectral preprocessing. The results showed that the coefficients of variation (CV) of total nitrogen (TN) and total phosphorus (TP) of the original and compound samples were reduced by 0.103 and 0.107, respectively, compared to the original samples. Both the homogeneity of concentration distribution and richness of spectral information were improved. Compared to the original model, the coefficient of determination (R²_pred) of TN and TP was promoted by 0.049 and 0.061, respectively. The residual predictive deviation (RPD) was improved by 1.547 and 0.176, respectively. Compared to the compound model, R²_pred of TN and TP was promoted by 0.026 and 0.022, respectively. RPD was improved by 0.470 and 0.052, respectively. The validation results showed that the R²_pred of TN model and TP model were 0.903 and 0.878 while RPD were 2.916 and 2.508, respectively. Researches indicate that not only the sample representativeness from calibration sets but also the prediction performance of models are availably improved by means of samples compounding, supporting the fast and accurate quantification of nutrients before the land application.