基于连续小波变换、SHAP和XGBoost的土壤有机质含量高光谱反演

针对土壤有机质含量高光谱反演中存在的光谱有效信号薄弱和光谱信息冗余问题，提出结合连续小波变换、SHAP和XGBoost的土壤有机质含量高光谱反演框架.以北京市延庆区和房山区永久基本农田土壤为研究对象，首先，基于连续小波变换处理的土壤光谱反射率数据构建初始XGBoost模型；然后，利用SHAP方法分析模型中各波段的贡献度以筛选特征波段；最后，基于特征波段重新构建和优化XGBoost模型，实现土壤有机质含量高光谱反演.发现连续小波变换尺度为25时，利用SHAP方法选取的40个特征波段构建的XGBoost模型准确性最高，有机质含量反演值和实测值之间的R²为0.80,RMSE为3.60g·kg^-1；随着连续小波变换尺度的增大，R2呈现先升高后降低的趋势，25尺度下的R2比21尺度的高0.37;SHAP方法选取的特征波段比Pearson相关分析法少682个，RMSE低0.69 g·kg^-1;XGBoost模型的R2分别比随机森林和支持向量机模型高4%和8%.验证了结合连续小波变换、SHAP和XGBoost在土壤有机质含量高光谱反演...

Hyperspectral Inversion of Soil Organic Matter Content Based on Continuous Wavelet Transform, SHAP, and XGBoost

Aiming to address the problems of weak spectral signals and redundant spectral information existing in hyperspectral inversion of soil organic matter content, a hyperspectral inversion framework combining continuous wavelet transform, SHAP, and XGBoost was proposed. Taking the permanent basic farmland soil in Yanqing District and Fangshan District of Beijing as the research object, an initial XGBoost model was first constructed based on the spectral reflectance data of soil processed by continuous wavelet transform. Then, the SHAP method was used to analyze the contribution of each band in the model to select the characteristic bands. Finally, the XGBoost model was reconstructed and optimized based on the characteristic bands to realize the hyperspectral inversion of soil organic matter content. It was found that the XGBoost model based on the 40 characteristic bands of continuous wavelet transform at the 2⁵ scale selected by the SHAP method showed the highest accuracy, with the R² and RMSE between the inversed and measured value of the organic matter content being 0.80 and 3.60 g·kg^-1, respectively. The R² first increased and then decreased with the increase in the scale of continuous wavelet transform, and the R² at the 2⁵ scale was 0.37 higher than that at the 2¹ scale. The number of characteristic bands selected by the SHAP method was 682 less than that by the Pearson correlation analysis method, and the RMSE was 0.69 g·kg^-1 lower. The R² of the XGBoost model was 4% and 8% higher than that of the random forest and support vector machine models, respectively. The results demonstrated the effectiveness of the combination of continuous wavelet transform, SHAP, and XGBoost in the hyperspectral inversion of soil organic matter content, which provides technical support for rapid and accurate monitoring of soil organic matter content.