Abstract:

【Objective】 To study rapid estimation of soil organic carbon content using hyperspectral data in the hope of providing scientific basis for rational development of land resources in lakeside oases in arid regions. 【Methods】 The north lakeside oasis of Bosten Lake was taken as the study area and the measured soil organic carbon content data were combined with the hyperspectral data. Successive Projections Algorithm (SPA) was used to screen the successive bands after SG smoothing (SG), Continuum Removal (CR) and Continuous Wavelet Transform (CWT) pre-processing for the original spectra. Support Vector Machines (SVM) models were used to estimate soil organic carbon content. 【Results】 Soil organic carbon content in the study area ranged from 0.69 g/kg to 50.32 g/kg, with an average value of 14.15 g/kg and a standard deviation of 9.51 g/kg, showing moderate variability and coefficient of variation of 67.20%. The original spectral reflectance of soil changed with the increase of wavelength, at 350-750 nm, the spectral reflectance increased, at 750-2,150 nm, the spectral reflectance showed a relatively stable trend; from 2,150 nm to 2,500 nm, the spectral reflectance gradually decreased. With the increase of decomposition scale, the local characteristics of the original spectrum of soil after pretreatment by continuous wavelet transform became more and more obvious, and the absorption and reflection peaks were becoming smoother and smoother. The feature bands selected by the continuous projection algorithm were concentrated in 350-952, 1,007-1,742 and 2,082-2,381 nm, and the feature bands only accounted for 0.30% of the Vis-NIR spectrum. The training set and verification set of the SVM model constructed by continuous wavelet transform and continuous projection algorithm were R²=0.76, RMSE=4.78 and R²=0.94, RMSE=3.30, RPD=2.50, respectively. 【Conclusion】 The CWT-SPA-SVM could be effectively estimate soil organic carbon content in the study area.

Key words: soil organic carbon content; continuous wavelet transform; successive projections algorithm; support vector machines model; hyperspectral data

摘要：

【目的】利用高光谱数据快速估算土壤有机碳含量,为干旱区湖滨绿洲合理开发土地资源提供科学依据。【方法】以新疆博斯腾湖北岸湖滨绿洲为研究区,将实测的土壤有机碳含量数据与高光谱数据相结合,对原始光谱进行SG平滑(SavitzkyGolay smoothing,SG)、连续统去除(Continuum Removal,CR)、连续小波变换(Continuous Wavelet Transform,CWT)预处理,采用连续投影算法(Successive Projections Algorithm,SPA)筛选特征波段;应用支持向量机(Support Vector Machines,SVM)模型估算土壤有机碳含量。【结果】(1)研究区土壤有机碳含量为0.69~50.32 g/kg,平均值为14.15 g/kg,标准差为9.51 g/kg,呈中等变异性,变异系数为67.20%。(2)土壤原始光谱反射率在350~750 nm,光谱反射率呈上升趋势,在750~2 150 nm,光谱反射率呈相对平稳趋势,在2 150~2 500 nm,光谱反射率逐渐下降;连续小波变换对土壤原始光谱预处理后随着分解尺度的增加,光谱局部特征明显,吸收峰和反射峰越来越平滑;采用连续投影算法筛选的光谱特征波段集中于350~952 nm、1 007~1 742 nm、2 082~2 381 nm,且特征波段仅占可见光-近红外光谱波段的0.30%。(3)连续小波变换结合连续投影算法构建的SVM模型,其训练集和验证集分别R²=0.76,RMSE=4.78和R²=0.94,RMSE=3.30,RPD=2.50。【结论】CWT-SPA-SVM可有效估算研究区土壤有机碳含量。

关键词: 土壤有机碳含量, 连续小波变换, 连续投影算法, 支持向量机模型, 高光谱数据