近红外光谱法快速测定土壤碱解氮、速效磷和速效钾含量

运用偏最小二乘法(PLS)和人工神经网络(ANN)方法分别建立了0.9 mm筛分风干黑土土壤碱解氮、速效磷和速效钾含量预测的近红外光谱(NIRS)分析模型。使用偏最小二乘算法建立的碱解氮、速效磷和速效钾校正模型的决定系数R²分别为0.9520、0.8714和0.7300，平均相对误差分别为3.42%、13.40%和7.40%。人工神经网络方法建立的碱解氮、速效磷和速效钾校正模型的决定系数分别为0.9563、0.9493和0.9522，相对误差分别为2.67%、6.48%和2.27%，测试集仿真的相对误差分别为5.44%、16.65%和7.87%。结果表明，人工神经网络方法所建立的校正模型均优于偏最小二乘法所建模型；用近红外光谱分析法预测土壤碱解氮含量是可行的，而速效磷、速效钾模型的测试集样品仿真的相对误差较大，其预测可行性还需做进一步研究。

Rapid prediction of available N, P and K content in soil using near-infrared reflectance spectroscopy

The calibration models were established using Partial Least Squares(PLS) and Artificial neural network(ANN) techniques to relate NIR spectral data to the concentrations of available N, available P and available K in 0.9 mm dried soil. Coefficients of determination(R²) between results from chemical analysis and NIR-predicted concentrations, based on calibrations of PLS, are 0.9520 for available N, 0.8714 for available P and 0.7300 for available K, and the mean relative errors of PLS models are 3.42%, 13.40% and 7.40%, respectively. Coefficients of determination, based on calibrations of ANN, are 0.9563 for available N, 0.9493 for available P and 0.9522 for available K, the mean relative errors of ANN models are 2.67%, 6.48% and 2.27%, respectively, and the mean relative errors of test samples are 5.44%, 16.65% and 7.87%, respectively. The results show that ANN technique is better than PLS in NIRS analysis, and the NIRS method is feasible to predict the concentration of available N, but the mean relative errors of test samples for available P and available K are high relatively, therefore, further study should be done in this field.