Enhancing Coastal Sea Level Predictions: A Hybrid Approach Combining TimeGAN-Augmented Data, and CNN-GRU Models

Sea level rise induced by climate change poses a significant threat to coastal cities, many of which serve as major economic hubs due to their strategic coastal locations, such as New York and Shanghai. To mitigate the increasing risks, accurate predictive models are urgently required to assess potential impacts effectively. In this study, we treat sea level data as signal data, leveraging its temporal structure to apply advanced signal processing techniques. Specifically, we employ Seasonal-Trend Decomposition using Loess (STL) to isolate the underlying trend by removing seasonal components and noise. This extracted trend is then used to train predictive models. We propose a hybrid framework that integrates STL decomposition with deep learning architectures, focusing on CNN-LSTM and CNN-GRU networks to capture both spatial and temporal dependencies in sea level data. To evaluate the performance of our models, we select three major coastal cities: Shanghai, New York, and Lisbon. Our empirical results demonstrate that augmenting training datasets with TimeGAN-generated data significantly enhances model performance. For ConvLSTM models, TimeGAN reduces the average mean squared error (AMSE) by approximately 66.1%, 76.6%, and 64.5% for Shanghai, New York, and Lisbon, respectively. Similarly, for ConvGRU models, TimeGAN achieves AMSE reductions of about 56.7%, 64.0%, and 63.3% for the same cities. In New York, ConvGRU (n = 0) achieves an AMSE of 0.522, outperforming ConvLSTM (n = 0), which records an AMSE of 1.348. Similarly, in Lisbon, ConvGRU (n = 0) achieves a significantly lower AMSE of 0.365 compared to ConvLSTM’s 0.512. In Shanghai, ConvGRU (n = 0) yields an AMSE of 0.364, slightly higher than ConvLSTM’s 0.311. These results reveal two key insights: (1) TimeGAN-generated data enhances prediction reliability across CNN-RNN architectures, and (2) CNN-GRU (n = 0) models consistently outperform their CNN-LSTM (n = 0) counterparts. Overall, this study demonstrates the effectiveness of treating sea level data as signal data and leveraging signal processing techniques to enhance predictive modeling, underscoring the vital role of such methods in addressing the challenges of climate change.