Estimating Soil-Water Characteristic Curve (SWCC) Using Machine Learning and Soil Micro-Porosity Analysis

Abstract

Abstract The soil water characteristic curve (SWCC) reveals soil porosity and soil-water interactions at different matric suctions. Numerous methods, such as laboratory determination, CT-scan, image analysis, and predictive models, have been employed to investigate soil porosity system and their correlation with the SWCC. Image analysis techniques offer valuable insight into soil pore system, providing data that cannot be obtained by other methods. This study aims to compare the role of image analysis technique in reconstructing the SWCC with the laboratory measurement method. Eight machine learning models and algorithms, including Gradient Boosting (GB), Ada Boost (AB), Decision Tree (DT), Random Forest (RF), Artificial Neural Network (ANN), Support Vector Machine (SVM), K-Nearest Neighborhood (KNN), and Linear Regression (LR), were utilized for the reconstruction of the SWCC using the Orange-3 data mining software. The predicted SWCCs by models were compared with the measured SWCC. The models used to reconstruction of SWCC were categorized as capable and incapable to SWCC prediction. Four statistical parameters, namely RMSE, MAE, Willmott’s index of agreement (d1), and R2, were utilized to assess the performance of the models using all input data. The GB, AB, DT were the top best models in correct prediction of SWCC. Among them the GB model achieved near-perfect predictions, with RMSE, MAE, d1, and R2 values of 0.016, 0.011, 0.94, and 0.982, respectively. However, when the matric suction was used as only input data, the performance of all the models declined. This study demonstrates that Orange-3 is user friendly software to predict SWCC, without labor preprocessing activities.