Machine learning-based hindcasting of significant wave heights in the Arabian Gulf

Abstract

This study developed a machine learning (ML) framework for long-term hindcasting of significant wave height (H S ) in the Arabian Gulf using Random Forest Regression (RFR) and eXtreme Gradient Boosting (XGB). The models were trained and tested using ERA5 wind and wave data at eight locations consisting of three deep water locations across the main axis of the Gulf and 5 nearshore locations along the Qatar coast considering 30, 20 and 10 years of training windows and subsequent 15, 10 and 5 years of testing windows, respectively. Performance evaluation and independent validation of model H S with ERA5 H S and Satellite-derived H S , respectively, reveals that both models performed reasonably well and remained stable across different temporal windows, indicating that the 10-year training period is good enough to capture the signals required for long-term and short-term variability in H S . Spatial transferability experiments showed that models trained at one representative location are well adequate to predict H S at any random locations in the Arabian Gulf. The study further confirms that the developed ML models are suitable for short-term and long-term predictions H S as it reproduces the diurnal variability and peaks associated with strong wind events with minimal errors.