Pedotransfer functions are valuable in hydrologic analysis because they transform readily available measurements into structured data. This work develops a pseudocontinuous pedotransfer function for prediction of cumulative infiltration using Soil Water Infiltration Global (SWIG) database. Ten different datasets were provided as input data of infiltration characteristics. The largest input data has more than 3250 infiltration curves and 80,000 datapoints involving classic and non-classic infiltration curves. This study is focused on the impact of the regularization technique and Artificial Neural Network (ANN)-based ensemble modeling on the accuracy of pedotransfer infiltration functions. To this end, the Multi-Layer Perceptron (MLP) was equipped with a Double-layer ELastic net regularization technique (i.e., MLP-DEL) for examining input selection (parsimony) and optimizing the geometry (sparsity) of the neural networks. A novel hybrid grey wolf optimizer-trust region algorithm is presented for the MLP-DEL. K-means and a self-organizing map network were embedded within the Bootstrap AGGregatING (Bagging) algorithm to investigate the role of ensemble modeling on infiltration pedotransfer functions. This paper’s results show that the optimal architecture of the single non-regularized ANNs for each of the ten input datasets generally consists of the Bayesian Regularization Backpropagation (BRB) with deep hidden layer and Tansig/Logsig activation functions. The investigation of sparsity-parsimony of the ANN pedotransfer models by MLP-DEL revealed that the regularization algorithm improves the accuracy of the single non-regularized ANN by removing the outliers provided that the ANN is not overly complex. The distribution of regularization factors in the two-stage regularization algorithm implied that while the accuracy of the ANN pedotransfer functions is dependent on all inputs, and the improvement of the accuracy occurs by re-arraignment of the hidden-output weights of the network. It is shown by this paper’s results that bagging ensemble modeling improves the RMSE, MAE and NSE of simulated values by 33%, 45% and 90%, respectively. It is also shown that the ensemble model compensates the weakened based learners such that an improvement of 198% was observed in the NSE of the ensemble bagging model compared to the single model.

中文翻译：

---

用于渗透曲线的分层伪连续机器学习 pedotransfer 模型：正则化和集成建模作用的研究


Pedotransfer 函数在水文分析中很有价值，因为它们可以将现成的测量值转换为结构化数据。这项工作开发了一种伪连续 pedotransfer 函数，用于使用土壤水渗透全球 （SWIG） 数据库预测累积渗透。提供了 10 个不同的数据集作为渗透特征的输入数据。最大的输入数据有超过 3250 条渗透曲线和 80,000 个涉及经典和非经典渗透曲线的数据点。本研究的重点是正则化技术和基于人工神经网络 （ANN） 的集成建模对 pedotransfer 渗透函数准确性的影响。为此，多层感知器 （MLP） 配备了双层 ELastic 网络正则化技术（即 MLP-DEL），用于检查输入选择（精简）和优化神经网络的几何形状（稀疏性）。提出了一种用于 MLP-DEL 的新型混合灰狼优化器-信任域算法。K-means 和自组织映射网络嵌入到 Bootstrap AGGregatING （Bagging） 算法中，以研究集成建模对渗透 pedotransfer 函数的作用。本文的结果表明，10个输入数据集中每个数据集的单个非正则化人工神经网络的最佳架构通常由具有深隐藏层的贝叶斯正则化反向传播（BRB）和Tansig/Logsig激活函数组成。MLP-DEL 对 ANN pedotransfer 模型的稀疏性-简约性的调查表明，如果 ANN 不太复杂，则正则化算法通过去除异常值来提高单个非正则化 ANN 的准确性。 两阶段正则化算法中正则化因子的分布意味着，虽然 ANN pedotransfer 函数的精度取决于所有输入，但精度的提高是通过重新排列网络的隐藏输出权重来实现的。本文的结果表明，装袋集成建模将模拟值的 RMSE、MAE 和 NSE 分别提高了 33%、45% 和 90%。研究还表明，集成模型补偿了较弱的基础学习者，因此与单个模型相比，集成装袋模型的 NSE 提高了 198%。