Coral sand cays are key transitional landforms that enhance reef resilience under climate change. Predicting their morphological evolution is essential for understanding formation processes. However, conventional models, relying on statistical extrapolation or numerical simulations with simplified assumptions, are limited in capturing the nonlinear and dynamic processes of cay evolution. Existing deep learning approaches extract spatiotemporal patterns more effectively but remain limited to profile-based analyses, lacking cay-wide, pixel-level predictions. Because cay evolution is influenced by marine environmental factors, integrating environmental forcings in a physics-aware manner is important for robust prediction, yet this remains underexplored. This study proposes a physics-aware deep learning model, MDF-ConvLSTM, which embeds oceanographic forcings into ConvLSTM units via a dual-level conditional gating mechanism to accurately predict cay evolution. SHapley Additive exPlanations (SHAP) is employed to quantify the magnitude and direction of variable contributions. Applied to Feixin Island, MDF-ConvLSTM outperforms CNN-LSTM and other benchmarks under both single- and multi-horizon scenarios, achieving Dice scores of 0.915 and 0.924, and IoU of 0.844 and 0.859, respectively. SHAP analysis indicates that the eastward wind component accounts for approximately 37% of the total feature importance. This study provides a novel model for predicting coral sand cay evolution, supporting reef management and conservation.
Zhang et al. (Tue,) studied this question.