论文部分内容阅读
建模方法是影响可见-近红外光谱定量结果的主要因素之一。在470～1000nm波段的12个土壤剖面对48个剖面样经过风干、研磨、过筛后进行光谱采集。经一阶微分变换及Savizky-Golay平滑处理后,分别应用主成分回归(PCR)、偏最小二乘回归(PLSR)和反向传播神经网络(BPNN)3种方法建立土壤全氮(TN)的定量模型。PCR与PLSR两线性模型的决定系数(R2)分别为0.74和0.8,其剩余预测偏差(RPD)分别为2.23和2.22,但两模型仅能用于TN的粗略估计。由PCR提供主成分数,PLSR提供潜变量(LV)数分别作为BPNN的输入所构建的两个非线性模型均明显优于线性模型PCR和PLSR。其中以4个LV作为输入的BPNN-LV模型预测性能最优,R2以及RPD分别达到0.9和3.11。实验结果表明,提取可见-近红外光谱的PLSR LV因子作为BPNN的输入,所建定量模型可用于土壤氮纵向时空分布的快速准确预测。
The modeling method is one of the main factors affecting the quantitative results of visible-near-infrared spectroscopy. 48 profile samples from 12 soil profiles in the 470-1000 nm band were air-dried, ground and sieved for spectral acquisition. After first-order differential transformation and Savizky-Golay smoothing, three methods of principal component regression (PCR), partial least squares regression (PLSR) and back propagation neural network (BPNN) were applied to establish soil total nitrogen Quantitative model. The determination coefficients (R2) of the two linear models of PCR and PLSR were 0.74 and 0.8, respectively, and the residual predictive bias (RPD) was 2.23 and 2.22, respectively, but the two models could only be used for the rough estimation of TN. The two non-linear models constructed by PCR with principal components and the number of latent variables (LVs) provided by PLSR as input to BPNN were significantly better than linear model PCR and PLSR. Among them, BPNN-LV model with 4 LVs as input has the best prediction performance, with R2 and RPD reaching 0.9 and 3.11 respectively. The experimental results show that the PLSR LV factor of visible-near-infrared spectroscopy can be used as the input of BPNN. The quantitative model can be used for the rapid and accurate prediction of soil longitudinal-temporal distribution of nitrogen.