Abstract
Coastal water-level information is essential for coastal zone management, navigation, and oceanographic research. However, long-term water-level observations are usually only available at a limited number of locations. This study discusses a complementary and simple neural network (NN) approach, to predict water levels at a specified coastal site from the data gathered at other nearby or remote permanent stations. A simple three-layer, feed-forward, back-propagation network and a neural network ensemble, named Atlantic Canadian Coastal Water Level Neural Network (ACCSLENNT) models, was developed to correlate the nonlinear relationship of sea level data among stations by learning from their historical characteristics. Instantaneous hourly observations of water level from five stations along the coast of Atlantic Canada—Argentia, Belledune, Halifax, North Sydney, and St. John’s—are used to formulate and validate the ACCSLENNT models. Qualitative and quantitative comparisons of the network output with target observations showed that despite significant changes in sea level amplitudes and phases in the study area, appropriately trained NN models could provide accurate and robust long-term predictions of both tidal and nontidal (tide subtracted) water levels when only short-term data are available. The robust results indicate that the NN models in conjunction with limited permanent stations are able to supplement long-term historical water-level data along the Atlantic Canadian coast. Because field data collection is usually expensive, the ACCSLENNT models provide a cost-effective alternative to obtain long-term data along Atlantic Canada.
Corresponding author address: Guoqi Han, Fisheries and Oceans Canada, St. John’s, NF A1C5X1, Canada. Email: guoq.han@dfo-mpo.gc.ca
