基于近红外光声光谱的土壤有机质含量定量建模方法

摘要

Near-infrared reflectance photoacoustic spectroscopy (NIRS) was applied to fast determination of soil organic matter (SOM) by different quantitative methods. The dried samples of five different soil types in the middle and eastern China were selected to assess the quantitative estimation of SOM based on different modeling methods. First derivative, multiplicative scatter correction (MSC) and smoothing algorithms were used to preprocess the original spectra before the calibration models were developed. The results showed that different pre-processing algorithms markedly affected the accuracy of SOM calibration models. The sequence of SOM models with different pre-processing algorithms was MSC+ Norris-gap first derivative smoothing filter (NGFD) > MSC >Norris first derivative smoothing filter > standard normal variate (SNV) > Norris-gap second derivative smoothing filter (NGSD) > LOG > Savitzky-Golay first derivative (SGFD) > Savitzky-Golay second derivative (SGSD). The spectra processed with the combination of MSC and NGFD performed the best among all the pre-processing algorithms. In addition, the calibration model based on PLS-BPNN displayed the highest estimation accuracy with R2 of 0.97 and RMSEP of 1.88, followed by PLS, SMLR and PCR, while the validation model with independent data gave R2 of 0.97 and RMSEP of 1.72, respectively. These results indicated that PLS-BPNN based on MSC-NGFD spectra was a potentially optimal method for SOM estimation.%该研究的目的在于应用近红外光声光谱技术结合不同的定量分析方法实现5种不同类型土壤有机质含量的快速估测.对中国中、东部地区5种不同类型土壤风干样本进行光谱扫描,经过多元散射校正、一阶导数、二阶导数及平滑等预处理后,应用逐步多元回归(SMLR)、主成分分析(PCR)、偏最小二乘法(PLS)和偏最小二乘法-反向传播神经网络(PLS-BPNN)等方法建立土壤有机质含量的定量估测模型.结果显示,不同预处理方法对所建土壤有机质含量估测模型的预测精度有较大影响,总体表现为多元散射校正+Norris一阶导数＞多元散射校正＞Norris一阶导数＞标准正态化＞Norris二阶导数＞吸光度＞Savitzky-Golay平滑后一阶导数＞Savitzky-Golay平滑后二阶导数.对于4种不同建模方法,均以多元散射校正+Norris一阶导数滤波平滑后的光谱建模精度最高,其中采用PLS-BPNN方法建模效果最好,其次是PLS、SMLR和PCR.采用PLS-BPNN建立有机质校正模型具有极高的预测精度,建模决定系数和均方根偏差分别为0.97和1.88,模型测试决定系数和均方根偏差分别为0.97和1.72.因此,基于多元散射校正+Norris一阶导数光谱建立的PLS-BPNN模型可能是土壤有机质含量估测建模的最优方法.