基于GWR模型的土壤重金属含量高光谱预测

Hyperspectral Prediction of Soil Heavy Metal Content Based on GWR Model

摘要:

针对地理加权回归分析（GWR）自变量选取问题，本文通过相关分析确定表征土壤重金属Cu含量与光谱特征关系的最佳光谱变换方式，利用普通最小二乘法建立候选光谱变量子集，运用回归分析法构建GWR模型，并将预测结果与传统方法构建的GWR模型进行对比。结果表明：对数微分变换光谱与土壤重金属Cu含量相关系数为-0.626~0.618，最能表征研究区域土壤光谱特征与土壤重金属Cu含量的关系，为研究区域最佳光谱变换方式。研究区土壤重金属Cu含量GWR预测最优模型的自变量组合为2 308、1 400、2 197、2 138 nm 4个波长，其构成GWR模型的赤池信息准则（AICc）最小，决定系数R^2和调整决定系数R^2_adj最大。与传统方法构建的GWR模型相比，本文方法构建的模型更简洁、稳健性更强、预测精度更高。

Abstract:

Aiming at the selection of independent variables in geographically weighted regression (GWR) analysis, the construction of GWR model was explored by using soil heavy metal Cu content as the dependent variable and soil hyperspectral data as the independent variables. The best spectral transformation that characterized the relationship between the soil heavy metal Cu content and the spectral reflectance in the study area was determined, throughing the correlation analysis of the spectral reflectance of the original soil and its eight transformation methods. The candidate variable subsets was established by using the ordinary least squares method, and the GWR model was constructed by using the regression analysis method. The advantages of the GWR model, as well as its prediction accuracy, were compared with that of the GWR model constructed by traditional methods. The results showed that the correlation coefficient between the logarithmic differential transformation spectrum and the soil heavy metal Cu content was from -0.626 to 0.618, which could best represent the relationship characteristics between the soil spectrum and the soil heavy metal Cu content in the study area, and it was the best spectral transformation method for the study area. The results of the regression analysis indicated that the optimal GWR prediction model for the soil heavy metal Cu content in the study area had four independent variables combined of 2 308 nm, 1 400 nm, 2 197 nm, and 2 138 nm wavelengths, which had the smallest Akaike information criterion (AICc), the largest determination coefficient (R^2) and adjust coefficient of determination (R^2_adj). Compared with the GWR model constructed by the traditional method, the GWR model constructed by the regression analysis method was more concise with stronger model robustness and higher prediction accuracy.