Dissertation/Thesis Abstract

Modelling Spatio-Temporal Relationships With Deep Neural Networks to Estimate Coastal Water Levels
by Yin, Lun, Ph.D., Stevens Institute of Technology, 2019, 239; 13857808
Abstract (Summary)

Coastal water level data are valuable in many monitoring, emergency management, forecast and research applications, yet observation gaps pose a challenge. This study uses multilayer perceptron and autoencoder-decoder models to learn the spatio-temporal relationships among water levels at 30 stations to estimate the missing water level data. The autoencoder approach is found to be the best to provide both accurate and stable estimations. With quality-controlled inputs, the autoencoder models achieve RMSEs ranging from 2.4 to 7.4 cm on out-of-sample data. The performances are substantially better than the results of Inverse Distance Weighting, which simply defines the spatial relationships as distance-based weights. Missing inputs, a critical issue left out of prior studies, are handled in this paper by the Designated Inverse Dropout method, which ignores the missing inputs and uses the remaining valid inputs to guarantee an output, and the symphony method, which replaces the missing inputs with model estimations at the other stations. With the symphony method of applying these models, the RMSEs are further reduced to between 2.2 and 6.5 cm, even outperforming the well-validated hydrodynamic model hindcasts from the Stevens Flood Advisory System which have RMSEs ranging from 4.2 to 11.3 cm. The resulting models have many applications beyond improving historical observations, including providing nowcast data to support real-time water surface mapping and data assimilation in operational hydrodynamic models, and establishing virtual stations to continue to provide water level data after a physical observation station is removed.

Indexing (document details)
Advisor: Orton, Philip
Commitee: Creamer, German, Datla, Raju, Hajj, Muhammad
School: Stevens Institute of Technology
Department: Schaefer School of Engineering & Science
School Location: United States -- New Jersey
Source: DAI-B 80/11(E), Dissertation Abstracts International
Subjects: Ocean engineering
Keywords: Autoencoder, Missing input, Neural networks, Surge prediction, Water level, Water level prediction
Publication Number: 13857808
ISBN: 978-1-392-24024-3
Copyright © 2020 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy