基于CWT-Stacking集成学习模型的灌区土壤水盐含量和pH值估算

Estimation of Soil Water-Salt Content and pH Value in Irrigation Districts Based on CWT-Stacking Ensemble Learning Model

摘要:

土壤含水率、含盐量、pH值精确估算对于干旱半干旱农业灌区的水土环境监测与可持续土地利用至关重要。本文以河套灌区农田和其他植被覆盖区为研究对象，采用一阶导数、二阶导数、包络线去除、标准正态变换以及连续小波变换对原始光谱进行预处理与特征增强，并选取偏最小二乘回归、支持向量机、随机森林、极端梯度提升和轻量级梯度提升机5种基学习器，结合Stacking、Voting、Bagging 3种集成策略，构建土壤水盐含量、pH值综合估算模型。结果表明：小波变换在去除高频噪声与凸显局部特征方面表现最优，适用性评分达到4.7，显著提升了光谱数据与土壤水盐含量、pH值的敏感性。相较于单一模型，集成学习策略有效降低了过拟合风险，并提高了模型泛化能力，其中Stacking集成策略综合表现最佳，对土壤含水率、含盐量、pH值训练集决定系数分别提升4.9%、4.7%和5.1%。基于CWT-Stacking构建的最优模型实现了目标水盐含量、pH值的高精度估算，土壤含水率、含盐量、pH值验证集决定系数分别达到0.85、0.89和0.82，均方根误差分别为2.26%、0.61 g/kg和0.30。先进的光谱变换技术与复杂的集成学习算法相结合，为干旱和半干旱地区大尺度土壤健康状况动态监测提供了高精度、可行的技术手段。

Abstract:

Accurate estimation of soil moisture, salinity and pH value is essential for soil and water environment monitoring and sustainable land use in arid and semi-arid agricultural irrigation districts. Taking farmland and other vegetation-covered areas in the Hetao Irrigation District as the research objects, the first-order derivative, second-order derivative, continuum removal, standard normal variate and continuous wavelet transform (CWT) were adopted to preprocess raw spectra and enhance features. Five base learners, including PLSR, SVM, RF, XGBoost and LightGBM were selected, and three ensemble strategies of Stacking, Voting and Bagging were combined to construct comprehensive estimation models for soil moisture, salinity and pH value. The results showed that continuous wavelet transform performed the best in removing high-frequency noise and highlighting local features, with an applicability score of 4.7, which greatly improved the sensitivity of spectral data to soil moisture, salinity and pH value. Compared with single models, ensemble learning strategies effectively reduced the risk of overfitting and enhanced the generalization ability of models. Among all strategies, Stacking achieved the best overall performance, and the coefficients of determination R2 of the training sets for soil moisture, salinity and pH value were increased by 4.9%, 4.7% and 5.1%, respectively. The optimal model constructed by CWT-Stacking realized high-precision estimation of soil moisture, salinity and pH value. The R2 of the validation sets reached 0.85, 0.89 and 0.82, and the corresponding root mean square errors (RMSE) were 2.26%, 0.61 g/kg and 0.30, respectively. The combination of advanced spectral transformation technology and complex ensemble learning algorithms can provide a high-precision and feasible technical means for the dynamic monitoring of large-scale soil health in arid and semi-arid regions.