土壤反射光谱综合反映了土壤的理化性质和内部结构, 高光谱遥感已被用于基于土壤反射光谱特性的土壤分类。 已有研究一般利用土壤反射光谱一阶微分主成分作为输入量进行光谱分类模型构建, 但主成分数据缺乏物理意义, 且缺乏对比性、 适用范围也有限。 与反射率一阶微分数据相比, 基于去包络线提取具有明确物理意义的特征参数, 能够提高土壤分类的精度, 并寻找到一种高精度土壤分类模型。 选取吉林省农安县的四种典型土壤(风砂土、 草甸土、 黑土、 黑钙土), 将采集后的土壤样本进行风干、 研磨、 过2 mm筛处理, 采用ASD FiledSpec^®3便携式光谱仪对处理后的土壤样本的可见光近红外光谱区进行测试, 从而获得土壤样本的光谱数据。 对光谱数据进行九点平滑、 10 nm重采样处理进行降噪, 将处理后的数据分别进行一阶微分主成分以及去包络线处理。 利用土壤样本的去包络线提取光谱特征参数。 以一阶微分主成分数据和光谱特征参数为输入量分别代入Logistic聚类模型(LR)、 人工神经网络聚类模型(ANN)、 K-均值聚类模型(K-means)。 首先明确了不同土类之间的反射光谱曲线、 去包络线的差异大小, 以及相同土壤的反射率曲线、 去包络线进行土壤分类的优劣, 并且在去包络线的基础上提取能够区分不同土类的光谱特征参数; 其次, 比较一阶微分主成分与光谱特征参数作为输入量时, 三种光谱分类模型精度差异并分析不同模型精度差异的原因。 结果表明: (1)四种土壤的反射光谱曲线差异较小, 去包络线可以极大的增强四种土壤在430~1 210 nm之间的光谱差异, 并在去包络线的基础上构建具有明确物理意义的光谱特征参数。 (2)将一阶微分主成分和光谱特征参数分别代入三种聚类模型可知, 以光谱特征参数为输入量的土壤光谱分类模型均超过了以一阶微分主成分为输入量的模型精度, 由于光谱特征参数保留了原数据的物理意义、 更准确的体现了不同土壤类型之间的差异性, 而一阶微分主成分数据带有一定的模糊性不同范围之间缺乏对比性, 在土壤分类中以光谱特征参数作为输入量更具有优势。 (3)在三类土壤分类模型中, LR的分类精度最高为76.67%, Kappa系数为0.56; ANN的分类精度中等为72.50%, Kappa系数为0.48; K-means的分类精度最低, 只有65.00%, Kappa系数为0.33。 研究成果可为土壤精细制图、 以及土壤分类仪器的研制提供技术支持。

The soil reflectance spectrum curve reflects the physical and chemical properties and internal structure of the soil. Hyperspectral remote sensing technology has been used to classify soil based on the soil reflectance spectrum characteristics. The first order differential principal component of soil reflectance spectrum is generally used to construct the spectral classification model, but the principal component data is lack of physical significance, contrast and limited scope of application. Compared with the first-order differential reflectivity data, the extraction of the characteristic parameters based on the de-enveloping line can improve the accuracy of soil classification and find a high-precision soil classification model. In this study, four typical soils (wind-sand soil, meadow soil, calcareous soil) were selected in Nong’an County, Jilin Province. The collected soil samples were dried, ground and treated by 2mm sieve. ASD FiledSpec^®3 portable spectrometer was used to measure the visible near infrared spectrum of the treated soil samples, and the spectral data of the soil samples were obtained. The spectral data were smoothed by nine points, the noise was reduced by 10nm resampling, and the processed data were processed by the first order differential principal component and the de-enveloping line respectively. The spectral characteristic parameters were extracted by using the continuum removed line of soil samples. The first order differential principal component data and spectral characteristic parameters were input into Logistic clustering model, artificial neural network clustering model and K-means clustering model respectively. In this paper, the reflectance spectra of different soils, the difference of the envelope, the reflectivity curve of the same soil, and the advantages and disadvantages of the soil classification are determined. And the spectral characteristic parameters which can distinguish different soil types are extracted on the basis of de-enveloping line. Secondly, when the first order differential principal component is compared with the spectral characteristic parameter as input, the accuracy differences of the three spectral classification models are compared and the reasons for the difference in the accuracy of different models are analyzed. The results showed that: (1) The difference of the reflectance spectra of the four soils was small, and the spectral difference between the four soils could be greatly enhanced by the continuum removedline. The spectral characteristic parameters with clear physical meaning are constructed on the basis of the de-enveloping line. (2) The first order differential principal component and spectral characteristic parameters are introduced into the three clustering models respectively. The soil spectral classification model with spectral characteristic parameters as input is more accurate than that of the first order differential principal component model, because the spectral characteristic parameters retain the physical meaning of the original data. More accurately reflects the differences between different soil types, and due to the fact that the first order differential principal component data have a certain degree of fuzziness and are lack of contrast between different ranges, it is more advantageous to use spectral characteristic parameters as input in soil classification. (3) Among the three soil classification models, the Logistic clustering model has the highest classification accuracy of 76.67% kappa coefficient of 0.56; the average classification accuracy of the artificial neural network model is 72.50% and the Kappa coefficient is 0.48 K-mean clustering model has the lowest classification accuracy, only 65.00% . And Kappa coefficient is 0.33. The research results can provide technical support for fine mapping of soil and the development of soil classification instrument.