Empowering Coastal Resilience: A Multi-Layer Perceptron Approach for Subseasonal-to-Seasonal Sea Level Predictions in the Gulf of Mexico

Subsidence and the impact of climate change are dual drivers that significantly contribute to the ongoing rise in sea levels and increased flood risk. At this juncture, the primary tool available to stakeholders for predicting these events is harmonic forecasting, which, while valuable, offers only short-term predictions spanning from a few hours to a couple of days. While short-term predictions are advantageous, they offer limited time for stakeholders' to prepare and implement necessary mitigation measures. The integration of Subseasonal-to-Seasonal (S2S) predictions could substantially enhance stakeholder preparedness. By affording methods to years of foresight, S2S predictions would empower stakeholders to proactively strategize and execute robust mitigation approaches. The true potential of S2S lies in its real-time operational capability, facilitating prompt decision-making. This research endeavors to actualize this potential by proposing the implementation of a machine learning model that can offer predictions spanning from 3 months up to 3-years.

Our research focuses on the Gulf of Mexico, a region of paramount importance to the US economy as it hosts a majority of the country’s major ports. This region is also significantly affected by subsidence, which markedly amplifies the incidence of coastal inundations. Presently, stakeholders grapple with approximating water level trends over the coming months and years using interannual variability, yielding a wide range of +/- 15 cm in Texas. Unfortunately, this range proves inadequate for informed decision-making, leaving stakeholders uncertain about the magnitude and direction of water level changes.

To address these limitations, we propose the implementation of a Multi-Layer Perceptron (MLP) model, utilizing a shared water level signal from various stations in Texas. This signal, derived from a distributed station strategy in the Gulf of Mexico, enables the creation of a model capable of generalizing predictions across the region. Recognizing the challenges of real-time modeling, arising from potential interruptions in input data, we have streamlined the inputs to focus solely on historical Texas water level signal measurements.

Preliminary findings showcase a promising mean absolute error of 7 cm when forecasting water level magnitudes spanning from 3 months to 3 years. This represents a substantial advancement compared to the existing interannual variability of 30 cm, which stakeholders currently rely upon. Importantly, our approach not only quantifies water level magnitudes but also discerns their directional trends, a feat currently unattainable. Consequently, the proposed operational S2S model stands as a pivotal asset for stakeholders and coastal managers, ensuring the safety and security of Texas communities along the coastline.