Integration of causal deep learning for tide level forecasting in data-scarce coasts

Abstract

Accurate prediction of tide gauge heights is crucial for effective coastal risk management, especially in data-scarce regions like Oman, where monitoring networks are limited. High Tide Flooding (HTF) behavior on Masirah Island is investigated in this study by combining meteorological records and a cutting-edge causal deep learning model. We compare four models—Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), its variant Bidirectional LSTM (Bi-LSTM), Temporal Convolutional Network (TCN), and a newly developed Causal-Bi-LSTM—for their suitability in modeling tide gauge levels using precipitation, temperature, wind speed, and sea level pressure derived from ERA5-Land climate variables. Our study reveals that HTF events are highly seasonal, with 70 % occurring in June, primarily due to the combined effects of tidal (median contribution: 53.24 %) and non-tidal residual (110.57 %) effects. The Causal-Bi-LSTM model performed best (Nash–Sutcliffe Efficiency = 0.961, Correlation Coefficient = 0.99) when incorporating physical constraints, while sensitivity analysis revealed that sea level pressure was the key driver. Future projections under SSP scenarios unveil alarming trends: while SSP126 maintains conditions relatively stable, SSP245 and SSP585 exhibit intensification and seasonality in HTF events, with the high emission scenario (SSP585) projecting 36 events by mid-century and extending vulnerability to months not affected before. These results highlight the need for region-specific adaptation strategies and reveal how causal deep learning can enhance flood forecasting in terms of both accuracy and physical interpretability. The study provides a methodology for coastal vulnerability assessment that can be applied to other data-scarce regions with rising flood hazards under climate change.