A neural network approach to predict soil salinity

Artificial neural networks (ANNs) have established their importance in situ ations where the input-output relationship of the system under investigatio n is either not explicit or is too complicated to be expressed by mathemati cal expressions. In this study, the potential of ANN models to predict deve lopment of soil salinity is explored. Experimental data were collected from a field lysimeter study on the build-up of salts under subirrigation with brackish water. Different architectures of ANN models with varying numbers of processing elements and hidden layers were explored. Also, different dat a preprocessing techniques were employed to account for the usual variabili ty in experimental data and for unbiased selection of data for ANN model de velopment. Statistical parameters, such as correlation coefficients, arithm etic mean, average absolute deviation, and standard deviation values illust rate the applicability of ANN models to predict soil profile salinity.