Comparative study of multivariate hybrid neural networks for global sea level prediction through 2050

Abstract

The ongoing rise in global sea levels poses significant risks to coastal regions such as storms surges, floodings and necessitates accurate predictive models to inform the relevant government organizations that are responsible of mitigation strategies. This study leverages advanced hybrid deep learning techniques to forecast global sea level changes up to the year 2050. Utilizing a combination of Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, our model integrates historical global sea level data from climate.gov and global air temperature projections from the CMIP6 (Coupled Model Intercomparison Project Phase 6) model. Performance evaluation, based on metrics such as Nash-Sutcliffe Efficiency, Mean Squared Error (MSE), and the Diebold-Mariano Test, demonstrates the superior accuracy of the hybrid models over traditional deep learning models. Results show that the hybrid LSTM-CNN model outperforms the standalone models, achieving an MSE of 0.4644 mm and an NSE of 0.9994, compared to the LSTM model’s MSE of 2.4450 mm and NSE of 0.9970. These findings underscore the potential of deep learning methodologies in enhancing the precision of long-term sea level predictions, providing valuable insights for policymakers and researchers in climate science.