Application of artificial neural networks for predicting watershed runoff

Predicting watershed runoff is complicated because of spatial heterogeneity exhibited by various physical and geomorphological properties that influence water movement. The goal of the present study was to explore innovative applications of artificial neural networks (ANNs) for predicting watershed runoff. ANNs were initially utilized for predicting monthly runoff over three medium-sized watersheds in Kansas. Conventional ANNs, and recurrent neural networks (RNNs) were constructed and assessed against other empirical approaches. Monthly precipitation and temperature formed the inputs, and monthly runoff was chosen as the output. Direct use of feedforward neural networks without time-delayed inputs did not provide a significant improvement over other regression techniques. However, inclusion of feedback with RNNs generally resulted in better performance. Modular neural networks (MNNs) were developed to address different rules associated with high, low, and medium runoff events for the three Kansas watersheds. Different modules within the network were trained to learn subsets of the input space corresponding to these events in an expert fashion. A gating network was used to mediate the response of all the experts. The problem was posed as one of Bayesian statistics combined with maximum likelihood estimation of network parameters. Three modular network architectures were examined: two based on hard classification, and one based on soft classification. In addition, a fully-connected feedforward network was developed for comparison purposes. Based on the results, modular networks were the preferred choice. Geomorphologic artificial neural networks (GANNs) were developed to explore the possibility of mapping the watershed geomorphology onto the network architecture. For such ANNs, the number of nodes in the hidden layer was equal to the number of possible paths for water particles to move to the watershed outlet, dictated by watershed geomorphology. The connection weights between hidden and output layers were specified by the path probability based on the theory of geomorphologic instantaneous unit hydrograph. The weights between input and hidden layers were free parameters that were trained on observed data. The trained networks were subsequently used to synthesize watershed runoff. GANNs were applied to the Back Creek and Indian-Kentuck Creek watersheds in Indiana, and were capable of simulating direct runoff hydrograph