基于高光谱特征参数优选的土壤盐分含量建模及其验证

目的 研究不同维度光谱变换下土壤盐分反演模型及其验证。方法 以博斯腾湖西岸湖滨绿洲为研究区,面向ASD高光谱数据,利用17种一维数学变换光谱和3种二维变换光谱指数,分别与实测土壤盐分进行相关分析,得到0.01显著性检验水平下初步优选的光谱特征参数,基于VIP准则选入最佳自变量实现PLSR模型构建,进行精度验证。结果 研究区干季土壤平均反射率随含盐量的增加而高于湿季土壤平均反射率,尤其体现在590、800、1 810、2 150 nm处;17种一维单波段光谱变换中,对数倒数的一阶微分(1/lgR)变换与土壤盐分含量相关性最好,峰值敏感波段为1 083 nm,相关系数绝对值|r|最高达0.63;3种二维两波段光谱变换中,归一化光谱指数NDSI(R_{1 780},R_{1 742})与土壤盐分含量相关性最好,相关分析决定系数R ²最大值为0.57;基于特征归一化光谱指数结合VIP准则进行自变量筛选的PLSR估算模型效果最佳,土壤盐分建模集和验证集的决定系数 {R_{\mathrm{V}}}^2 达0.77,均方根误差RMSEV为0.64 g/kg,相对分析误差RPD为2.11。 结论 利用归一化光谱指数NDSI建立PLSR高光谱模型可有效地对研究区土壤盐分进行定量估算。

Modeling and Verification of Soil Salt Content Based on Hyperspectral Characteristic Parameter Optimization

【Objective】 To explore the feasibility of soil salinity inversion model under different dimensional spectral transformations.【Methods】 The lakeside oasis on the west of Bosten Lake in Xinjiang was taken as the research area, the correlation analysis between 17 one-dimensional mathematical transformation spectra and 3 two-dimensional transformation spectral indexes of ASD hyperspectral data and the measured soil salinity were conducted to obtain the preliminarily optimized spectral characteristic parameters at the significance test level of 0.01. Then, the PLSR model was constructed based on the VIP criteria and selected into the optimal independent variable and the accuracy was verified.【Results】 The average reflectance of dry soil was higher than that of wet season with the increase of salt content, especially at 590, 800, 1,810 nm and 2,150 nm. Among the 17 one-dimensional single-band spectral transformations, the first derivative of logarithmic reciprocal (1/lgR) had the best correlation with soil salinity, the peak sensitive band was 1083nm, and the absolute value of correlation coefficient was up to 0.63. Among the three two-dimensional two-band spectral transforms, the normalized spectral index NDSI(R_{1 780}, R_{1 742}) had the best correlation with soil salt content, and the maximum value of correlation analysis determination coefficient R ² was 0.57. The PLSR estimation model based on characteristic normalized spectral index and VIP criterion for independent variable screening had the best effect. The determination coefficient {R_{\mathrm{V}}}^2 of soil salt modeling set and verification set was 0.77, the root mean square error was 0.64 g/kg, and the relative analysis error was 2.11.【Conclusion】 Using the normalized spectral index (NDSI) to establish PLSR hyperspectral model could effectively estimate the soil salinity in the study area.