Comparative performance of multiple linear regression and artificial neural network models in estimating solute-transport parameters

Abstract

Indirect estimate of solute-transport parameters through pedo-transfer functions (PTFs) is becoming important due to expensive and time-consuming direct measurement of the parameters for a large number of soils and solutes. This study evaluated the relative performance of PTFs of multiple linear regression (MLR) and Artificial Neural Network (ANN) models in predicting velocity (<em>V</em>), dispersion coefficient (<em>D</em>) and retardation factor (<em>R</em>) of CaCl<sub>2</sub>, NaAsO<sub>2</sub>, Cd(NO<sub>3</sub>)<sub>2</sub>, Pb(NO<sub>3</sub>)<sub>2</sub> and C<sub>9</sub>H<sub>9</sub>N<sub>3</sub>O<sub>2</sub> (carbendazim) in five agricultural soils. <em>V</em>, <em>D</em> and <em>R</em> of the solutes were determined in repacked soil columns under steady-state unsaturated water flow conditions. Textural class, particle size distribution, bulk density, organic carbon, relative pH, clay%, grain size, and uniformity coefficient of the soils were determined. MLR and ANN models were calibrated with the measured data of four soils and verified for another soil. Root-Mean Square Error (RMSE) is significantly smaller (0.015) and modelling efficiency (EF) is significantly larger (0.999) for ANN model than those (0.096 and 0.954, respectively) for MLR model. Negative Mean Absolute Error (MAE) (-0.0002) of MLR model indicates overestimation, while positive MAE (0.00003) of ANN model indicates minimal underestimation. The ANN model is less biased than the MLR model during prediction. Thus, the ANN model can significantly enhance pollution transport prediction through soils with good accuracy.