基于智能学习的光滑土壤表面光谱反射率模型

An Intelligent Learning-Based Model of Spectral Reflectance for Smooth Soil Surfaces

摘要: 土壤光谱反射率是遥感反演的关键参数。本研究采集东北地区700个干土样品的实验室光谱数据，基于SHAP方法分析土壤理化性质及成土母质等协变量的贡献，构建了基于梯度提升回归(GBR)的干土光谱反射率模型(400~2500 nm)，并结合比尔–郎伯定律预测湿土反射率。结果表明：1) 干土模型预测精度较高(R² = 0.93, RMSE = 0.018)，引入成土母质使精度提升了4.5%；2) 模型预测的反射率空间分布与卫星观测结果具有良好一致性；3) 湿土反射率模拟精度在8%、16%、20%、24%和30%五个质量含水量梯度下，R²分别达到0.78、0.85、0.85、0.89和0.85。综上，本研究提出了一种高效且具有良好可解释性的土壤光谱反射率建模方法，可为土壤反射率预测及基于遥感的土壤性质反演提供可靠支撑。

Abstract: Soil spectral reflectance is a critical parameter for the remote sensing retrieval of land surface properties. In this study, laboratory spectral data were collected from 700 dry soil samples across Northeast China. The contributions of environmental covariates, including soil physicochemical properties and parent material, were analyzed using the SHAP method. Subsequently, a dry soil spectral reflectance model spanning 400~2500 nm was developed based on Gradient Boosting Regression (GBR), and wet soil reflectance was predicted by coupling this model with the Beer-Lambert law. The results indicate that: 1) The dry soil model exhibited high predictive accuracy (R² = 0.93, RMSE = 0.018), and the incorporation of parent material improved the model’s accuracy by 4.5%. 2) The spatial distribution of the model-simulated reflectance demonstrated good consistency with satellite observations. 3) Under five gravimetric water content gradients (8%, 16%, 20%, 24%, and 30%), the simulation accuracy (R²) for wet soil reflectance reached 0.78, 0.85, 0.85, 0.89, and 0.85, respectively. In conclusion, this study proposes an efficient and highly interpretable modeling methodology for soil spectral reflectance, providing reliable support for reflectance prediction and the remote sensing-based retrieval of soil properties.