Two-Stage Machine Learning Methodology for prediction of morphological changes along coastline

Kumar, Pavitra ORCID: https://orcid.org/0000-0002-4683-724X and Leonardi, Nicoletta (2024) Two-Stage Machine Learning Methodology for prediction of morphological changes along coastline. In: AGU Fall Meeting 2024, 9 December 2024 - 13 December 2024, Washington, D.C., USA. (Unpublished)

Accepted Version File not available for download. Available under License In Copyright. Download (116kB)

Abstract

Coastal waves continuously reshape the coastline, a process accelerated by the effects of climate change. As coastlines change rapidly, it is crucial to understand and predict coastal changes to protect coastal communities and assets. By analyzing field data from beach transects along the Morecambe Bay coastline, this study observed significant erosion at most of these beaches since 2007. Data was collected from 2007 to 2022, twice a year, at 125 locations along the coastline. Over more than a decade, the analysis revealed that the maximum mudflat retreat exceeded 2 km since 2007. Additionally, the maximum marsh retreat was approximately 500 m, and clifftop retreat was about 5 m. To further model erosion and predict future changes in Morecambe Bay, this study utilized machine learning models capable of learning non-linear relationships and accurately forecasting future data. A two-stage modeling methodology was adopted. In the first stage, a random forest classifier was developed to categorize beach behavior into four types: eroding, accreting, stable, or experiencing short-term fluctuations. The second stage involved using LSTM and sequence-to-sequence models, which took the first stage's output to predict the available sediment volume after erosion or accretion. The LSTM model achieved a testing regression score exceeding 0.9 for one-step-ahead (6 months) predictions of sediment volume time series. Similarly, the sequence-to-sequence model achieved a testing regression score exceeding 0.9 for both three-time-step-ahead (1.5 years) predictions and ten-time-step-ahead (5 years) predictions.