基于高光谱反演的复垦区土壤重金属含量经验模型优选

Empirical Model Optimization of Hyperspectral Inversion of Heavy Metal Content in Reclamation Area

摘要:

以工矿复垦区为实验区域，基于ASD FieldSpec 4高光谱遥感数据，结合实测的土壤重金属含量，利用回归分析与特征选择方法，开展了基于高光谱数据的土壤重金属含量反演研究与实验并进行了经验模型优选。通过对光谱曲线进行一阶微分、对数一阶微分以及对数倒数的一阶微分等数学变换有效提高了光谱数据与土壤重金属含量的相关性。在此基础上采用偏最小二乘回归（Partial least squares regression, PLSR）、随机森林回归（Random forest regression, RFR）、支持向量机回归（Support vector machine regression, SVMR）3种回归分析模型开展土壤重金属含量反演实验，结果表明偏最小二乘回归（PLSR）对研究区内土壤中重金属含量的反演最为有效，尤其对区域内主要障碍因子镉（Cd）元素含量的反演效果最佳，验证集决定系数R2为0.76。基于粒子群算法（Particle swarm optimization, PSO）、遗传算法（Genetic algorithm, GA）、Relief F算法 3种特征选择方法对偏最小二乘回归（PLSR）模型进行优化，结果表明粒子群算法（PSO）可有效降低特征波段变量维度，进一步提高模型反演精度，使决定系数R2由0.76提高至0.84。综上，基于高光谱数据，采用偏最小二乘回归（PLSR）与粒子群算法（PSO）相结合的方法，可有效对工矿复垦区土壤中的重金属含量进行测度，可为复垦区土地的质量和生态指标监测提供理论方法和技术支持。

Abstract:

Taking industrial and mining reclamation land as the research object, based on the ASD FieldSpec 4 hyperspectral remote sensing data, combined with the field survey data of soil heavy metal attributes, using regression analysis and feature selection methods, the retrieval research and experiment of soil heavy metal content based on hyperspectral data were carried out, and the selection and comparison of empirical models were conducted. The correlation between soil heavy metal concentration and spectral data was effectively improved by the first derivative and logarithmic reciprocal etc. On this basis, three regression analysis models, including partial least squares regression (PLSR), random forest regression (RFR) and support vector machine regression (SVMR) were used to carry out the inversion experiment of heavy metal content in soil. The results showed that the partial least squares regression (PLSR) had the highest precision for the retrieval of heavy metal concentration in the reclaimed soil, especially for the cadmium (Cd) concentration, which was the main obstacle factor in the area. The determination coefficient (R2) of fit for the set was 0.76. Particle swarm optimization (PSO), genetic algorithm (GA) and Relief F were used to optimize the partial least squares regression (PLSR) model. The results indicated that PSO can effectively reduce the dimension of characteristic band variables and further improve the model inversion. And the R2 of fit was increased from 0.76 to 0.84. In conclusion, based on hyperspectral data, the combination of partial least squares regression (PLSR) and particle swarm optimization (PSO) can effectively measure the concentration of heavy metals in the soil of industrial and mining reclamation area, and it can provide theoretical methods and technical support for the detection of land quality and ecological indicators in the reclamation area.