Predicting Stream Nitrogen Concentration from Watershed Features Using Neural Networks

ÐThe present work describes the development and validation of an arti®cial neural network (ANN) for the purpose of estimating inorganic and total nitrogen concentrations. The ANN approach has been developed and tested using 927 nonpoint source watersheds studied for relationships between macro-drainage area characteristics and nutrient levels in streams. The ANN had eight independent input variables of watershed parameters (®ve on land use features, mean annual precipitation, animal unit density and mean stream ̄ow) and two dependent output variables (total and inorganic nitrogen concentrations in the stream). The predictive quality of ANN models was judged with ``hold-out'' validation procedures. After ANN learning with the training set of data, we obtained a correlation coecient r of about 0.85 in the testing set. Thus, ANNs are capable of learning the relationships between drainage area characteristics and nitrogen levels in streams, and show a high ability to predict from the new data set. On the basis of the sensitivity analyses we established the relationship between nitrogen concentration and the eight environmental variables. Key wordsÐneural network, back-propagation, modelling, nonpoint source pollution, nitrogen, watershed, land use, ecology