| 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】为了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】本文使用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】结果显示，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。另外本研究还评价了各模型的运行效率。【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： |
| 分类号:TV93 |
| 基金项目:三维非均质农田水力层析扫描研究与应用(51779179)；基于数据价值分析的根区土壤水分预测研究(51609173)；大数据环境下的土壤水及地下水均衡评估理论和方法研究(51861125202) |
|
| Comparison of Pedotransfer Function Models Based on Machine Learning Methods |
|
ZUO Bingxin1, ZHA Yuanyuan2
|
|
1.The School of Water Resources and Hydropower Engineering, Wuhan University;2.State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University
|
| Abstract: |
| 【Background】The accurate description and prediction of soil moisture state and movement process requires accurate soil hydraulic parameters. The PedoTransfer Functions (PTFs) can easily obtain soil hydraulic parameters with less time and effort. For different modeling methods, there are differences in model prediction accuracy and operating efficiency. 【Objective】In order to explore the advantages and disadvantages of different machine learning methods to establish pedotransfer function models under uniform conditions, 【Method】this study used the UNSODA database with global soil hydraulic properties to compare and analyze the prediction accuracy of artificial neural network (ANN), support vector machine (SVM), and K-nearest neighbor (KNN) on the parameters of van Genuchten model and the saturated hydraulic conductivity. All the model was trained and tested with the same soil samples, using sand, silt, clay percentage and bulk density as the input information to predict the parameters of van Genuchten model and the saturated hydraulic conductivity. We used the measured water content-matrix potential data and the saturated hydraulic conductivity of soil samples for the evaluating the performance of each model from the perspective of soil texture. 【Result】The results showed that the SVM model has the best predictive performance on the sample, followed by ANN, and the KNN model has a relatively lower predictive performance due to edge effects. In addition, this study also evaluated the operating efficiency of each model. The results show that when training the model, the ANN model takes the longest time, while the KNN model takes the shortest time. However, while during forecasting, the ANN model takes the shortest time, while the KNN model takes the longest time. In addition, we compared the popular used Rosetta model, which is a pedotransfer function model using a single hidden layer neural network. The results showed that the multi hidden layer neural network used in this study can improve the prediction accuracy of the model. In addition, using a more appropriate modeling method and increasing the amount of information of input data will help to improve the prediction accuracy of the model as well.【Conclusion】In this study, it is recommended to use the SVM model when building pedotransfer function models on small datasets since it can balance prediction accuracy and operational efficiency. However, the prediction efficiency of the ANN model is higher, and on large database, the ANN model can use mini-batch method to training the model without increasing or a small amount of calculation, so this research recommends that a more reasonable selection of modeling methods should be comprehensively considered in terms of calculation cost and prediction accuracy on large database. |
| Key words: pedotransfer function; machine learning; artificial neural network; support vector machine; K-nearest neighbor |
