我国北方寒地温差大, 土壤温差对近红外光谱测量土壤墒情有较大影响。 针对这一问题, 以北方寒地土壤为研究对象, 探究大范围温度胁迫下(-20~40 ℃)土壤的近红外光谱与土壤不同含水率之间的关系预测模型方法。 选取黑龙江八一农垦大学农学院试验基地中的黑土, 经烘干、 过筛等操作处理后配置含水率范围在15%~50%内八种不同湿度的土壤样品, 建立北方寒地土壤大范围温度胁迫下土壤的近红外光谱信息与含水率之间的定量预测模型。 在全波段光谱数据的基础上, 结合五种不同光谱信号预处理方法, 采用BP神经网络算法、 优化支持向量机算法(SVM)、 高斯过程算法(GP)三种智能算法建立北方寒地土壤近红外光谱与含水率的预测模型并验证模型的效果。 利用69组数据进行训练建模, BP神经网络相关参数设置为学习速率0.05, 最大训练次数设置为5 000, 隐层单元数确定为20; SVM采用径向基函数, 并利用leave-one-out cross validation确定了最佳惩罚参数为0.87, 使模型预测的准确性提高; 高斯过程算法内部采用马顿核。 模型的定量评估采用决定系数( R²)和均方根误差(RMSE)。 结果表明, 在建立的全部BP神经网络模型中, 效果最佳的为S_G-BP神经网络模型, 模型的 R²为0.960 9, RMSE为2.379 7; 在SVM模型中SNV-SVM模型的效果最好, 模型的 R²为0.991 1, RMSE为1.081 5; 在GP模型中S_G-GP模型的效果最好, 模型的 R²为0.928, RMSE为3.258 1, 综上基于SNV预处理的SVM模型训练效果最优。 利用剩余的35组光谱数据作为预测集验证模型性能, 经模型对比分析发现基于SVM算法的预测模型效果优于其他两种算法, 其中基于S_G的SVM模型效果最优, 其预测模型的 R²和差RMSE分别为0.992 1和0.736 9。 综合建模集与预测集的参数最终确定基于S_G的SVM模型为最佳模型。 此模型可以作为大范围温度胁迫条件下(寒地)的土壤含水率有效预测方法, 为设计优化适宜寒地便携式近红外土壤含水率快速测量仪提供科学依据。

There is a large temperature difference between summer and winter in northern China. Soil temperature difference greatly influences the measurement of soil moisture by NIR (Near-Infrared). A prediction model for soil NIR spectrum and soil moisture content under a wide range of temperature stress (-20~40 ℃) was introduced. Soil samples were collected in the experimental field of Heilongjiang Bayi Agricultural University. After drying and sieving, soil samples were dampened to moisture content ranging from 15% to 50%. Prediction model for NIR and soil moisture content under different temperature stress was built. 69 groups of spectral data was used as training set to build model based on the full-band spectral data and five different spectral signal preprocessing methods. BP (Back-propagation) neural network, optimized support vector machine (SVM) algorithm and Gaussian process algorithm (GP) were used to establish the prediction model of soil near-infrared spectrum and moisture content in northern cold areas,and verify the effect of the model. The learning rate for BP neural network was 0.05, the maximum training time was 5 000, and the number of hidden layer units was 20. SVM used the radial basis function and Leave-One-Out Cross-Validation to determine the optimal penalty parameter (0.87), which improved the accuracy of the model prediction. Marton kernel internally was used for the GP model. GP model was evaluated by the coefficient of determination ( R²), and root mean square error (RMSE). Results show that the S_G-BP neural network model has the best performance among the BP neural network models, with R² of 0.960 9 and RMSE of 2.379 7. The SNV-SVM model has the best performance among the SVM models with R² of 0.991 1 and RMSE of 1.081 5. The GP models, S_G-GP model has the best performance among GP models, with R² of 0.928 and RMSE of 3.258 1. In conclusion, the SVM model based on SNV preprocessing has the best training performance. 35 groups of spectral data were used as a prediction set to verify the model performance. According to the model comparison and analysis, the prediction model based on the SVM algorithm is better than the other two algorithms, among which the S_G-based SVM model has the best performance. R² and RMSE are 0.992 1 and 0.736 9, respectively. Combining the parameters of modeling set and prediction set, the SVM model based on S_G has the best performance in this study. This model can predict soil moisture content under a wide range of temperature stress in cold regions, providing a theoretical foundation for the design and optimization of portable NIR soil moisture rapid measurement instruments in the cold region.