基于成像光谱技术的土壤剖面发生层划分初探

土壤剖面是土壤发生学研究的核心，但过去几十年以来研究土壤剖面的技术并没有发生质的变化。成像光谱技术可以提供高空间、高光谱分辨率的土壤剖面数据，能够弥补反射光谱技术采样深度间隔较大的不足，用于定量研究土壤属性连续深度变化。以室内采集的土壤剖面成像光谱数据为研究对象，采用支持向量机方法进行光谱数据主成分分类，探讨成像光谱数据用于剖面发生层划分的可行性并分析影响因素。研究中首先定性分析各发生层平均光谱曲线形态特征，然后通过分析剖面光谱数据主成分深度变化特征及其散点分布情况，探讨其用于剖面发生层划分的可行性；最后进行1 000次随机划分数据集并建模、预测以减小误差，定量证明成像光谱数据用于土壤发生层划分的可行性，并通过样本分类错误频率来分析影响分类精度的因素。研究结果表明，受成土过程影响剖面内各发生层平均光谱曲线特征存在差异。成像光谱数据的主成分可以定量呈现土壤剖面深度方向上属性的连续变化及样本散点分布的集聚特征，能较好反映发生层之间的差异性，可以用于发生层划分。建模预测结果表明发生层的预测精度平均值达到93.08%。同时发现，光谱主成分分布相似区域的样本及位于发生层过渡区域的样本分类错误率较高。该研究为利用成像光谱技术进行土壤剖面发生层划分提供了理论依据，为下一步进行剖面发生层制图奠定了技术基础。

Horizon Classification in Soil Profile Using Imaging Spectroscopy

The soil profile is one of the core topics in pedogenesis research, but traditional pedological observations of soil profiles rely on the use of visible light and a toolbox that has not changed in the past decades. The imaging spectroscopy can provide high-resolution spatial and spectral soil profile information, which gives continuous depth functions of soil properties and compensates for the large gap between the sampling depths of reflectance spectroscopy. The objective of this study is to analyze the classification of soil horizon in a profile by investigating the spectral data of imaging spectroscopy collected in the laboratory. The support vector machine (SVM) method was used to classify the spectral data, and the feasibility and influence factors of the imaging spectroscopy for classification were studied. Firstly, the morphological characteristics of the average spectral curves of each horizon in sample profile were analyzed qualitatively. Secondly, depth dynamic and scatter plot of principal components were qualitatively analyzed to explain the feasibility of horizon classification using profile imaging spectroscopy. Finally, One thousand times computations were carried out to reduce the classification error by partitioning random dataset and building prediction model. The prediction results can quantificationally testify the feasibility and the influence factors were discussed by the percentage of wrong classification in prediction results. The results indicated that the four average spectral curves in sample profile differed and reflected the variation in depth derived from pedogenic processes. The principal components of the imaging spectral data showed the continuous change in the depth direction of the soil profile and the grouping feature in scatter plot, which proved that imaging spectroscopy reflected the difference between the genetic horizons and can be used for the horizon classification. The average accuracy of classification prediction reached 93.08%. Moreover, it was found that the sample with similar scattering distribution and locating transition region were classified in wrong classes easily. This study provides a theoretical basis for horizon classification, and proves that imaging spectroscopy is a potential technology for mapping soil profile.