TESTING AN ARTIFICIAL NEURAL-NETWORK FOR PREDICTING SOIL HYDRAULIC CONDUCTIVITY

Multilinear regression has been used extensively to predict soil hydra ulic properties, both the theta(h) and K(h) relationships, from easily obtainable soil variables. As an alternative, this study investigated the performance of an artificial radial basis neural network in predi cting some K(h) values from other variables. This kind of neural netwo rk may be seen as a multivariate interpolation technique, which can th eoretically fit any nonlinear continuous function. Neural networks are characterized by parameters that must be optimized to solve a given p roblem. We used a fitting procedure requiring only two parameters to e nsure a unique solution. These two parameters were determined by data splitting. Hypothetical data bases with uncertainties were simulated t o analyze the performance of the neural network in predicting a nonlin ear relation derived from a classical model for K(h). A soil data base covering a broad spectrum of soil horizons was used to test the neura l network in solving multivariate problems. Numerical simulations show ed that the neural network was sensitive to large uncertainties in the data base. It was more efficient than a multilinear regression when t he uncertainties were small. Experimental results showed that the neur al network was more efficient than the multilinear regression for pred icting K(h = -1 m) or K(h = -2.5 m) from two qualitative and five quan titative soil variables. It was also more efficient than two independe nt multilinear regressions, one for the sandy samples and the other fo r the loamy and clayey samples. Provided that a large data base with a ccurate K values is available, artificial neural networks can be usefu l to predict theta(h) and K(h) over a broad spectrum of soils.