| 摘要: |
| 土壤水分状态及运动过程的准确描述与预测需要精确的土壤水力参数，而使用土壤转换函数的方法可以方便省时省力地获得土壤水力参数。对于不同的建模方法，在模型预测精度、运行效率等方面有所差异。【目的】探究不同机器学习方法建立土壤转换函数模型的优劣。【方法】使用UNSODA全球土壤水力性质数据库，对比分析了人工神经网络（ANN）、支持向量机（SVM）、K-最近邻居法（KNN）对于van Genuchten模型参数的预测精度，并使用土壤样本的含水率-负压实测数据，从土壤质地的角度评价了各模型的预测效果。【结果】SVM模型对样本的预测效果最好，ANN次之，KNN模型受边缘效应影响预测效果稍逊。在训练模型时，ANN模型用时最长，KNN模型用时最短。然而，在预测过程中，ANN模型用时最短，KNN模型用时最长。【结论】本研究推荐在小数据集上建立土壤转换函数模型时使用SVM，而在更大型的数据集上要综合考虑计算成本、预测精度等方面合理选取建模方法。 |
| 关键词: 土壤转换函数；机器学习；人工神经网络；支持向量机；K-最近邻居法 |
| DOI：10.13522/j.cnki.ggps.2020445 |
| 分类号: |
| 基金项目: |
|
| Using Machine Learning Methods as a Pedotransfer Function to Estimate Soil Hydraulic Parameters |
|
ZUO Bingxin, ZHA Yuanyuan
|
|
State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University, Wuhan 430072, China
|
| Abstract: |
| 【Background】Accurate prediction of soil water dynamics requires soil hydraulic parameters and the PedoTransfer Function (PTF) is an indirect method estimating soil hydraulic parameters based on easy-to-measure soil properties.【Objective】The purpose of this paper is to compare different machine learning methods as a pedotransfer function to estimate soil hydraulic parameters. 【Method】We used the UNSODA soil hydraulic property database and compared three machine learning methods: artificial neural network (ANN), support vector machine (SVM), and K-nearest neighbor (KNN). The hydraulic parameters were described by the van Genuchten formula, and the relationship between its parameters with fractions of sand, silt and clay, as well as soil bulk density was analyzed using the three methods. The accuracy of each method was evaluated using measured water retention curves and saturated hydraulic conductivity from soils with different textures.【Result】The SVM model was most accurate and KNN the least to predict soil hydraulic parameters using these easy-to-measure soil properties. Evaluation of operating efficiency of all three methods revealed that the ANN model was least efficient and the KNN the most in model training. In contrast, the ANN model was most efficient while KNN the least in predicting soil hydraulic parameters. Comparison of the three models against the Rosetta model – a commonly used neural-network pedotransfer function with a single hidden layer – found that neural-network models with multiple hidden layers, as used in this paper, were more accurate. We also found that for all three models, increasing the number of input data improved their estimation accuracy.【Conclusion】Of the three models, SVM is most accurate for predicting soil hydraulic parameters using fractions of clay, sand and silt, and bulk density, followed by ANN, when the database was not large enough. With the size of the database increasing, the ANN model becomes increasingly more efficient. Since ANN can use the mini-batch method to train the model without increasing computational costs, our results suggest that selecting a suitable method to calculate soil hydraulic parameters should consider computational cost and estimation accuracy when the size of the database increases. |
| Key words: Pedotransfer function; machine learning; artificial neural network; support vector machine; K-nearest neighbor |
