Artificial neural network optimization to predict saturated hydraulic conductivity in arid and semi-arid regions

Saturated hydraulic conductivity (K-sat), one of the critical soil hydraulic properties, is used to model many soil hydrological processes. Measurement of K-sat on a routine basis is a labor-intensive, time-consuming, and expensive process. Alternatively, prediction of K-sat values from easy to obtain soil features is more economical and saves time. Artificial neural networks (ANNs) can be used to model and describe the most influential features affecting K-sat. This study aimed to develop and evaluate the potential use of generalized regression neural network (GRNN) to identify the optimal set of soil features to predict K-sat under arid and semi-arid environments. A total of 165 soil samples were collected from three depths (0-15, 15-30, and 30-60 cm) and analyzed for K-sat, texture, organic matter (OM), pH, bulk density (BD), and electrical conductivity (EC). Fourteen GRNN models were built with different feature combinations to identify the optimal set to predict K-sat. The results showed that soil texture explained 78% of the variability in soil K-sat while introducing EC improved model's ability to estimate soil Ksat (R = 0.93, MSE = 2.89 x 10(-12) m2 S-2). The optimum set of soil properties that should be included in the model were sand and clay percentages and EC values as evidenced from the cross-validation results. The GRNN model (using small dataset and set of features) provided reliable predictions of Ksat on bar with more complex models that included extensive set of features and used more extensive dataset. This work has implications for soil scients as provides an economical method to estimate K-sat values.

Automated summary

The generalized regression neural network (GRNN) can be used to predict the saturated hydraulic conductivity (K-sat) of soil, and soil texture and electrical conductivity can be used as the influential features in the prediction model. The GRNN model using a small dataset and limited features can provide reliable predictions of K-sat.