Prediction of watershed runoff using Bayesian concepts and modular neural networks

Artificial neural networks have been widely used as models for a variety of nonlinear hydrologic processes including that of predicting runoff over a watershed. In this paper such networks were organized in a modular architec ture to handle complex sets of rainfall-runoff data. Such data often contai n examples corresponding to different rules that may be associated with hig h, low, and medium streamflows. Different modules within the network were t rained to learn subsets of the input space in an expert fashion. A gating n etwork was used to mediate the response of all the experts. The problem was posed as one of Bayesian statistics combined with maximum likelihood estim ation of network parameters. The training of the gating network was equival ent to the classification problem (i.e., identification of the expert), whi le the experts trained to minimize the absolute difference between predicte d and target monthly discharges. The performance of modular networks in pre dicting runoff over three medium-sized watersheds was examined. Average mon thly rainfall of current and previous months and average monthly temperatur es were treated as network inputs, and monthly runoff was treated as output . Three different modular network architectures were examined in this payer : two based on hard classification and one based on soft classification. In addition, a fully connected feed forward network was utilized for comparis on purposes. On the basis of the results, modular networks appear to be goo d alternatives for predicting runoff. The role of these networks on issues regarding probabilistic interpretability and classification are discussed.