Use of imaging spectroscopy to assess different organic carbon fractions of agricultural soils
Use of imaging spectroscopy to assess different organic carbon fractions of agricultural soils
复制标题
使用成像光谱法评估农业土壤的不同有机碳成分
作者：
Michael Vohland;M. Harbich;O. Schmidt;T. Jarmer;C. Emmerling;S. Thiele
这项研究的地点位于德国莱茵兰-法拉蒂纳(“比特堡古特兰”)，涵盖了不同的地质底物和农业土壤区。总共在田间采集了42块地的样本；在实验室对顶层土壤样本进行了总有机碳(OC)、热水可提取碳(HWE-C)和微生物碳(CMIC)的分析。在地面战役的同时，2009年8月27日获得了HyMapTM机载成像传感器的数据集。经前处理后，用HyMap光谱分析测定OC、HWE-C和CMIC的含量。作为校准方法，我们使用了偏最小二乘回归(PLSR)，因为它允许处理大输入空间和噪声模式。由于HWE-C和CMIC的校正质量较差(交叉验证的R2值小于0.5)，我们还将PLSR与遗传算法(GA)相结合来预先选择一组最优的光谱特征，而不是使用全光谱。使用这种GA-PLSR方法，交叉验证中所有成分的结果都有相当大的改善(R2≥0.72)。所有碳组分非常相似的GA选择模式表明，虚假的(间接)相关性可能与评估HWE-C和CMIC有关。对于GA方法，不能排除由于基于C分数和光谱变量之间的机会相关性的选择而导致的一些过度拟合。
The site for this study - located in Rhineland-Palatinate, Germany ("Bitburger Gutland") - covered different geological substrates and agro-pedological zones. In total, 42 plots were sampled in the field; soil samples from the top horizon were analysed in the laboratory for total organic carbon (OC), hot water-extractable C (HWE-C) and microbial C (Cmic). In parallel to the ground campaign, a data set of the HyMapTM airborne imaging sensor was acquired on 27th of August 2009. After pre-processing, HyMap spectra were used to assess the contents of OC, HWE-C and Cmic. As calibration method we used partial least squares regression (PLSR), as it allows a handling of large input spaces and noisy patterns. Since calibration quality was poor for HWE-C and Cmic (cross-validated r2 values were less than 0.5), we additionally combined PLSR with a genetic algorithm (GA) to preselect an optimum set of spectral features instead of using the full spectrum. With this GA-PLSR approach, results improved considerably for all constituents in the crossvalidation (r2 ≥ 0.72). Very similar GA selection patterns for all carbon fractions suggest that spurious (indirect) correlations may be relevant for assessing HWE-C and Cmic. For the GA approach, some overfitting due to a selection based on chance correlations between C fractions and spectral variables cannot be excluded.