顾及粗糙度的土壤有机碳成像高光谱估测模型

可见近红外非成像光谱分析技术已被广泛用于土壤有机碳（SOC）含量估测，然而该技术的使用受土壤粗糙度的影响，对样本的前处理要求较高，导致模型的实用性受限。针对这一问题，以美国爱荷华州农田土壤为研究对象，使用成像及非成像光谱仪获取土壤样本研磨前后的可见近红外反射光谱，采用去包络线（CR）、吸光度变换（AB）、S-G平滑（SG）、标准正态变换（SNV）、多元散射校正（MSC）5种光谱预处理手段，利用偏最小二乘回归（PLSR）和支持向量回归（SVR）算法构建并对比土壤SOC光谱估算模型，探究利用成像光谱数据估测高粗糙度样本SOC含量的可行性。实验结果表明，使用成像光谱数据能够实现高粗糙度样本的SOC含量估算，而使用非成像光谱数据则无法估算高粗糙度样本的SOC含量；基于成像光谱数据建立的高粗糙度SOC最优PLSR估算模型R2能够达到0.739以及最优SVR估算模型R2为0.712，而基于非成像光谱数据建立的高粗糙度SOC最优PLSR和SVR估算模型R2仅仅分别为0.344和0.311。基于AB，SG，SNV和MSC这4种预处理手段之后的成像光谱数据建立的土壤样本研磨前的PLSR模型性能优于样本研磨之后建立的PLSR模型，而SVR模型性能正好相反。而对于非成像光谱数据来说，土壤样本研磨后建立PLSR和SVR模型精度总是强于样本研磨前建立的模型精度。对于这两种光谱数据和两个估算模型而言，不同的光谱预处理方法提高模型估算精度的能力不同。土壤样本研磨前后，基于成像光谱数据建立的PLSR和SVR模型性能均优于非成像光谱数据所构建的模型。成像光谱技术能够增强高粗糙度土壤样本可见近红外光谱与SOC的相关性，从而提高模型估算精度；能够克服土壤粗糙度的影响；为野外大尺度估测SOC含量提供了新的手段。

Estimation of Soil Organic Carbon Content by Imaging Spectroscopy With Soil Roughness

Visible and near-infrared (VIS-NIR) non-imaging spectroscopy has been widely applied to estimate soil organic carbon (SOC) content. Due to the high demand for soil sample pretreatments, VIS-NIR non-imaging spectroscopy easily suffers from soil roughness in practical application. This study explored the potential of imaging spectroscopy to estimate SOC content with high soil roughness. With soil samples collected in Iowa State, United States, imaging spectra were utilized to measure the VIS-NIR spectra of soil samples with and without ground. With five spectral pre-processing including continuum removed (CR), absorbance transformation (AB), S-G smoothing (SG), standard normal variate (SNV), and multiplicative scatter correction (MSC), partial least squares regression (PLSR) and support vector regression (SVR) were used to build estimation models to analyze the potential of imaging spectra. Non-imaging spectra were also applied to build PLSR and SVR models as a comparison. Results demonstrated that imaging spectra could achieve SOC content estimation for soil samples with high roughness, but non-imaging spectra could not successfully estimate that. The best PLSR and SVR model developed by imaging spectra could reach 0.739 and 0.712 of R2 for SOC content estimation of soil samples with high roughness, while that established by non-imaging spectra could achieve 0.344 and 0.311 of R2. Based on the imaging spectra after the four pre-processing methods of AB, SG, SNV, and MSC, the performance of the PLSR model established before soil sample grinding was better than that of the PLSR model established after soil sample grinding, while the performance of the SVR model was just the opposite. For non-imaging spectra, the accuracies of PLSR and SVR models established after soil samples grinding were always better than that of models established before soil samples grinding. For these two spectral data and the two estimation models, different spectral pre-processing methods had different abilities to improve the estimation accuracy of the model. The performance of imaging spectroscopy outperformed non-imaging spectra before or after being ground soil samples. Imaging spectra could enhance the correlation coefficient between VIS-NIR spectra and SOC for soil samples with high roughness, there by improving PLSR model’s performance. Our findings provide a new way to estimate SOC content on large-scale yield because imaging spectra could overcome the influence of soil roughness.