水稻根系生物量及碳氮比的可见-近红外光谱原位反演研究

The major sources of soil organic matter are crop straws and roots, whose biomass and ratio of carbon to nitrogen (i.e. C/N) are served as important driving parameters in many biogeochemical models for simulating soil organic carbon dynamics. Given the difficulties in obtaining roots in regional surveys, such as time-consuming in digging method and soil coring, little is known about the spatial distribution of crop roots and their C/N ratios, which is restricting the study on carbon and nitrogen cycles of cropland ecosystem. Therefore, paddy fields of Yujiang County located in the major area of rice production in China are selected as a subject for study. Firstly, in-situ visible near-infrared reflectance spectroscopy (Vis-NIRS) is collected from soil cores during the harvest period using the combination of a portable spectrometer and a plant contact probe. Then, these soil cores are sent to a laboratory to measure the biomass of roots and their C/N ratios. After different spectral preprocessing operations, the biomass and C/N ratio of rice roots are predicted by four approaches, namely, stepwise multiple linear regression, principal components regression, partial least square regression and back propagation neural network methods. Finally, the optimal spectral preprocessing and modeling approaches are confirmed based comprehensive comparisons of their fitting performance. This study hopes not only to be able to develop an approach for quickly and accurately monitoring in-situ biomass of roots and their C/N ratios, but also to serve the study on carbon and nitrogen cycles of cropland ecosystem at regional scales.

作物秸秆和根系是土壤有机质的主要来源，其生物量及碳氮比是许多生物地球化学模型的重要驱动参数。目前区域调查中通常使用挖掘法和土钻法采集作物根系，这些方法耗时费力，只适用于少量样点的观测，因而对全国或区域尺度上作物根系及碳氮比的空间分布知之甚少，严重制约了农田生态系统碳氮循环研究。基于土壤可见-近红外反射光谱对有机化合物的敏感性，本项目选择我国南方水稻主产区的江西省余江县作为研究区，在水稻收获期，先用土钻法采集土芯样品，再联合野外便携式地物光谱仪和植物接触探头，原位测定土芯光谱，然后将土芯样品带回实验室测定根系生物量及碳氮比。经不同光谱预处理后，采用多元逐步回归、主成分回归、偏最小二乘回归和BP神经网络法等方法建模，反演水稻根系生物量及碳氮比，并对其拟合效果进行比较，确定最佳的光谱预处理和建模方法，以期建立一种快速准确的地预测田间原位监测根系生物量及碳氮比的方法，服务于区域农田生态系。

作物秸秆和根系是土壤有机质的主要来源，其生物量及碳氮比是许多生物地球化学模型的重要驱动参数。目前区域调查中通常使用挖掘法和土钻法采集作物根系，这些方法耗时费力，只适用于少量样点的观测，因而对全国或区域尺度上作物根系及碳氮比的空间分布知之甚少，严重制约了农田生态系统碳氮循环研究。基于土壤可见-近红外反射光谱对有机化合物的敏感性，本项目选择我国南方水稻主产区的江西省余江县作为研究区，在水稻收获期，先用土钻法采集124个土芯样品，再联合野外便携式地物光谱仪和植物接触探头，然后测定土芯光谱、水稻根系生物量及土壤碳氮磷等含量。经不同光谱预处理后，最终采用偏最小二乘回归（PLSR）、主成分回归（PCR）、支持向量机回归（SVMR）和BP神经网络法（BPNN）等方法建模，预测水稻根系生物量及土壤碳氮磷等属性，并对其拟合效果进行比较，确定最佳的光谱预处理和建模方法。基于170个建模集，非线性SVMR模型（R2 = 0.90; RMSE = 3.92; Slope = 0.86）模拟水稻根系密度的性能明显明优于PCR和PLSR两种线性模型，而PCR（R2 = 0.83; RMSE = 5.16; Slope = 0.80）和PLSR（R2 = 0.85; RMSE = 4.95; Slope = 0.78）两种模型具有相似的模拟精度。相对于PCR和PLSR模型，SVMR模型能够降低21%–24%的均方根误差。类似地，对于四种土壤属性（SOM、TN、TP和TK）来说，采用非线性SVMR模型，SOM在建模集（R2CV = 0.90; RMSECV = 4.92; RPDCV = 3.08）和验证集（R2P = 0.88; RMSEP = 4.87; RPDP = 2.84）中的预测精度均最高，表明所建立的模型可对验证集样品作粗略估测。同样采用SVMR模型，TN在验证集（R2P = 0.86; RMSEP = 0.31; RPDP = 2.69）中取得了和SOM类似的预测精度。土壤TP和TK含量的预测结果相对较低，尤其是TK预测结果不理想。以上结果表明，SVMR模型与其它多变量模型相比，具有较大的优越性。SVMR模型能够避免高维光谱数据处理时必须面对的众多问题，较好地解决了非线性、高维数和局部极小点等问题，适合基于高光谱信息进行建模处理。