Forecasting discharge in Amazonia using artificial neural networks

The Amazon, located in northern South America, is the world's largest river basin, and covers an area of about 6.5 million km(2). The observed interan nual variability in precipitation and water availability during its main di scharge season has been shown to be influenced by Pacific and Atlantic Ocea n sea surface temperatures (SSTs). However, the links between large-scale a tmospheric motion and local and regional runoff patterns are essentially co mplex and still not fully understood. The processes involved are believed t o be highly non-linear, spatially and temporally variable, and not easily d escribed by physical or conceptual models. Artificial neural networks (NN) were trained to forecast discharge, one and two seasons in advance, at ten river sites in Amazonia from Pacific and At lantic Ocean SST anomalies. The NN with an input layer of eight neurons, on e hidden layer with 20 neurons and a one-neuron output layer was trained us ing back-propagation with momentum and gradient descendent. Results confirmed that different oceanic regions have distinct influences o n different parts of the Amazonian basin. Better forecasts for basins in th e northern part of Amazonia were obtained from Pacific Ocean SST and from A tlantic Ocean SST for basins in the southern part. Correlation coefficients between observed and estimated discharge (validation) were as high as 0.76 at some of the sites studied. The inclusion of precipitation as input impr oved the forecast for sites where NN did not perform well with training by SST only as input. The results obtained during this study corroborate and improve results obta ined previously by means of linear statistical methods. Copyright (C) 2000 Royal Meteorological Society.