不同粒径对土壤有机质含量可见—近红外光谱预测的影响

Effect of Different Particle Sizes on the Prediction of Soil Organic Matter Content by Visible-Near Infrared Spectroscopy

土壤有机质(SOM)是表征土壤肥力的重要指标,实现其快速准确检测可为精准农业区域管理提供有效的数据支撑。土壤粒径对SOM的光谱预测及仪器开发有很大的影响,为了明确不同粒径对SOM预测的影响,分别制备了1~2,0.5~1,0.25~0.5,0.1~0.25和<0.1mm五种均匀粒径及<1mm混合粒径共计6种粒径土样并进行了可见-近红外(300~2500nm)光谱数据采集。采用蒙特卡罗交叉验证分别剔除了不同粒径的异常样本,结合Savitzky-Golay卷积平滑法对光谱数据进行平滑去噪处理,比较了不同粒径样品的光谱反射率差异,并对平滑后的原始光谱R进行倒数IR、对数LR、一阶导数FDR等3种光谱变换并分析与SOM含量的相关性,基于竞争性自适应重加权算法(CARS)对光谱数据进行了特征波长提取,并结合偏最小二乘回归(PLSR)分别建立了相应的SOM含量预测模型。结果表明,不同粒径土样的平均光谱反射率与变异系数随着粒径的减小逐渐增加,且在大于540nm波长范围内,差异明显。随着粒径的减小,SOM含量与光谱反射率在全波段范围的相关性变化幅度愈加明显,FDR变换可明显改变全波段范围与SOM含量的相关性。通过CARS算法对FDR变换后的光谱数据进行特征波长提取,筛选出特征波长数为全波段数量的13.1%,降低了光谱数据重叠及无效信息干扰。对比不同SOM预测模型的结果,FDR变换光谱的建模精度较好,且粒径越小其模型的效果越好,特别在粒径<0.1mm时,模型的R达到0.91,RMSEP为2.20g·kg^(-1),RPD为3.33。基于CARS特征变量构建的SOM含量预测模型中,粒径<0.1mm预测模型的效果最好,R为0.78,RMSEP为3.00g·kg^(-1),RPD为2.00,可以实现SOM含量的可靠预测,且其他粒径下的模型仍有可优化的空间。该研究可以为实现SOM田间动态预测及仪器设计提供理论及模型参考。

Soil organic matter is an important indicator that characterizes soil fertility information,and realizing its rapid and accurate detection can provide effective data support for precision agriculture regional management.The particle size of the soil has a great influence on the spectrum prediction of SOM content and instrument development.To analyze the impact of different particle sizes on SOM prediction,five soil samples with the uniform particle size of 1~2,0.5~1,0.25~0.5,0.1~0.25,<0.1mm,and mixed particle sizes of<1mm were prepared,and the visible-near infrared(300~2500nm)spectral data was collected.Monte Carlo cross-validation was used to eliminate abnormal samples of different particle sizes,and the spectral data were smoothed and de-noised by the Savitzky-Golay convolution smoothing method.The spectral reflectance differences of samples with different particle sizes were compared,and three spectral transformations were performed on the smoothed original spectrum R,including reciprocal IR,logarithmic LR,and first derivative FDR.The correlation between SOM content and the reflectance of different transformed spectra was analyzed.The characteristic wavelength of the FDR transformed spectral data was extracted based on the Competitive Adaptive Reweighted Sampling(CARS)algorithm.Moreover,combined with the partial least squares regression(PLSR)to establish the corresponding prediction models of SOM content.The results show that the average spectral reflectance and coefficient of variation of soil samples with different particle sizes gradually increase with the decrease of particle size,and the difference is obvious in the wavelength range greater than 540nm.With the decrease in particle size,the correlation between SOM content,particle size,and spectral reflectance in the whole band range become more obvious.FDR transformation can significantly change the correlation between SOM content and spectral reflectance.The CARS algorithm was used to extract the characteristic wavelengths from the FDR transformed spectral data,and the number of characteristic wavelengths was screened out and reduced to 13.1%of the total number of bands,which reduced the overlap of spectral data and the interference of invalid information.Comparing the results of different SOM prediction models,the FDR transformed spectrum had good modeling accuracy.Especially when the particle size was less than 0.1mm,the model’s R~2,RMSEP and RPD value was 0.91,2.20g·kg^(-1),and 3.33.Among the SOM content prediction models constructed based on CARS characteristic variables,the prediction model with particle size<0.1mm has the best effect.Its R~2reached 0.78,RMSEP was3.00g·kg^(-1),and RPD was 2.00,which can achieve reliable prediction of SOM content,and there is still room for optimization of models under other particle sizes.This research can provide a reference for the rapid and accurate prediction of SOM content in the field environment and the design of instruments.