Artificial neural networks of soil erosion and runoff prediction at the plot scale

Neural networks may provide a user-friendly alternative or supplement to complex physically based models for soil erosion prediction for some study areas. The purpose of this study was to investigate the applicability of using neural networks to quantitatively predict soil loss from natural runoff plots. Data from 2879 erosion events from eight locations in the United States were used. Neural networks were developed for data from each individual site using only eight input parameters, and for the complete data set using 10 input parameters. Results indicated that the neural networks performed generally better than the WEPP model in predicting both event runoff volumes and soil loss amounts, with exception of some small events where the negative erosion predictions were not physically possible. Linear correlation coefficients (r) for the resulting predictions from the networks versus measured values were generally in the range of 0.7 to 0.9. Networks that predicted runoff and soil loss individually did not perform better than those that predicted both variables together. The type of transfer function and the number of neurons used within the neural network structure did not make a difference in the quality of the results. Soil loss was somewhat better predicted when values were processed using a natural logarithm transformation prior to network development. The results of this study suggest the possibility for using neural networks to estimate soil erosion by water at the plot scale for locations with sufficient data from prior erosion monitoring.