Artificial neural networks for estimating soil hydraulic parameters from dynamic flow experiments

Inverse methods are often used for estimating soil hydraulic parameters from experiments on flow of water through soil. We propose here an alternative method using neural networks. We teach a problem-adapted network of radial basis functions (RBF) the relationship between soil parameters and transient flow patterns using a numerical flow model. The trained RBF network accurately identifies soil parameters from flow patterns not contained in the training scenarios. A comparison with the inverse method (Annealing-Simplex) reveals a similarly good prediction by both approaches for randomly perturbed data and data from the real world. Nonetheless, the inverse method showed dependency on initial parameter estimates not required by the RBF network. Training demands moderately more computation and manpower than the inverse technique, but the absolutely stable and simple network application requires negligible resources. Thus, for individual applications, the network approach is slightly surpassed by the Annealing-Simplex method. However, the RBF network has to be trained only once and, subsequently, it can be applied easily and without effort upon any number of laboratory experiments with standardized experimental setups.