Inclusion of fractal dimension in machine learning models improves the prediction accuracy of hydraulic conductivity

Abstract

Measurement of hydraulic conductivity (HC) in the field and laboratory is time-consuming, laborious, and expensive, pedo-transfer functions can be used to predict the soil HC using easy-to-measure soil properties like bulk density (BD), soil texture, fractal dimension (D), organic carbon (OC) and glomalin content. In this study, 121 soil samples were used to predict HC using Multi Linear Regression, and four machine learning-based models i.e., Artificial Neural Network (ANN), Support Vector Machine (SVM), Classification and Regression Trees (CART) and Random Forest (RF). Two sets of input data were used i.e., dataset 1: texture data, BD, OC, and glomalin content and dataset 2: D, BD, OC, and glomalin content (Dataset 2). The models were evaluated based on Mean Absolute Error, Mean Absolute Percentage Error, Nash–Sutcliffe model efficiency, Root Mean Square Error (RMSE), and correlation coefficient. ANN with three hidden layers performed significantly for both input sets. The RMSE value was decreased by 17% in the training dataset and by 5.55% in the testing dataset when D was added to the input set for ANN. For both datasets, RF performed better and outperformed CART in predicting HC. According to the results, SVM with dataset 2 outperformed all other models which showed the inclusion of D in the dataset could predict HC more efficiently. However, further study is required for different combinations of datasets for evaluating the prediction efficiency of machine learning models for various regions.