Comparative Study of Statistical, Numerical and Machine Learning-based Pedotransfer Functions of Water Retention Curve with Particle Size Distribution Data

Abstract

The water retention curve (WRC) describes the nonlinear relation of soil water content (SWC) and matric potential. Since direct measurement of SWC and matric potential is difficult and time consuming, indirect approaches including statistical, numerical, and pattern recognition-based pedo-transfer functions (PTFs) that relate basic soil properties to the WRC have been developed during the last few decades. Although several studies reporting the performance of these models can be found in literature, it seems that an extensive investigation which compares the available models and introduces a reliable method to soil hydrologists can be useful. Therefore, in this study, the performance of multiple linear regressions (MLR) models, scaled numerical models and machine learning methods including artificial neural networks (ANN) and adaptive neuro-fuzzy inference systems (ANFIS) are compared using 98 UNSODA codes with various soil textures to estimate WRC. Results showed that regardless of the soil texture, ANN (RMSE = 0.029) predicts the WRC more accurately than ANFIS (RMSE = 0.035), scaled model (RMSE = 0.060) and MLR (RMSE = 0.071), respectively. Considering the soil texture, ANFIS performance is the best in the moderate and fine textured soils, while scaled numerical model predicts with acceptable performance in sandy soils. WRC prediction using easily available soil characteristics particularly when there is a lack of data, shows that newly developed machine learning methods are capable of predicting WRC considerably accurate for sustainable water flow and solute transport management.