Neural network input selection for hydrological forecasting affected by snowmelt

Snowmelt largely affects runoff in watersheds in Nordic countries. Neural networks (NN) are particularly attractive for streamflow forecasting whereas they rely at least on daily streamflow and precipitation observations. The selection of pertinent model inputs is a major concern in NNs implementation. This study investigates performance of auxiliary NN inputs that allow short-term streamflow forecasting without resorting to a deterministic snowmelt routine. A case study is presented for the Riviere des Anglais watershed (700 km(2)) located in Southern Quebec, Canada. Streamflow (Q), precipitations (rain R and snow S, or total P), temperature (T) and snow lying (A) observations, combined with climatic and snowmelt proxy data, including snowmelt flow (Q(SM)) obtained from a deterministic model, were tested. NN implemented with antecedent Q and R produced the largest gains in performance. Introducing increments of A and T to the NNs further improved the performance. Long-term averages, seasonal data, and Q(SM) failed to improve the networks.