区域不同土层土壤容重建模方法比较

Comparison of Modeling Methods for Region Soil Bulk Density in Different Soil Layers

【目的】构建土壤转换函数,利用土壤其它属性间接预测土壤容重。【方法】本文采用最优子集、lasso压缩估计、十折交叉检验等方法,以土壤有机质、土壤含水量、土壤质地和土壤采样深度作为预测变量,分不同垂直尺度(0~10、10~20、20~40、0~40 cm)对土壤容重进行预测。【结果】土壤容重系统变异的影响因子随土层中有机质含量高低的不同而不同;基于融合十折交叉检验的最优子集和lasso压缩估计所构建的土壤容重预测模型比已有模型精度更高(RMSE=0.063、0.029),但不同方法在不同垂直尺度上预测效果的表现不同,最优子集在表层0~20 cm所得土壤转换函数较lasso压缩估计效果更佳,而在土层20~40、0~40 cm,lasso压缩估计预测效果更为接近实际的土壤容重的分布趋势。【结论】与以往建模方法相比,十折交叉检验规避了在选取模型训练集和验证集上的随机性,提高了模型的预测精度;最优子集保证了全局最优的变量组合;lasso压缩估计弥补了最小二乘模型方差偏大的缺陷,使模型输出的稳健性得以提升,并刻画出模型的具体形式。研究成果可为区域土壤容重预测提供方法支撑。

【Objective】Soil conversion function was constructed to predict soil bulk density indirectly by other soil properties.【Method】In the present paper,with soil organic matter,soil water content,soil texture and soil sampling depth as predictors,soil bulk density was predicted in different vertical scales(0-10,10-20,20-40,0-40 cm)by optimal subset,lasso compression estimation,ten-fold cross-validation.【Result】The influence factors of soil bulk density system variation varied with the content of organic matter in soil layer;the prediction model of soil bulk density based on the optimal subset of fusion cross test and lasso compression estimation had higher accuracy than the existing models(RMSE=0.063,0.029),but the performance of different methods on different vertical scales was different,and the optimal subset was 0-20 cm in the surface layer.Compared with lasso compression estimation,the prediction effect of lasso compression estimation was better under 20-40 and 0-40 cm soil layers.【Conclusion】Compared with the previous modeling methods,the ten fold cross test avoids the randomness in the selection of model training set and verification set,improves the prediction accuracy of the model;the optimal subset ensures the global optimal combination of variables;lasso compression estimation makes up for the defect of large variance of the least squares model,improves the robustness of the model output,and describes the specific form of the model.The research results in the paper can provide support for modeling methods for regional soil bulk density prediction.