Optimizing neural network training for nearshore sea state forecasts using maximum dissimilarity algorithm

Accurate and timely wave forecasting is crucial for supporting vessel operations and reducing risks to coastal infrastructure and ecosystems. However, the high computational demands of traditional numerical models limit their applicability for real-time forecasting and nowcasting in nearshore environments. This study proposes an efficient alternative based on artificial neural networks (ANNs) trained with data selected via the maximum dissimilarity algorithm (MDA). By extracting representative subsets from long-term offshore wind and wave reanalysis data, the MDA drastically reduces the training dataset size while preserving the diversity of the wave climate. These subsets are propagated to shallow waters using a numerical model to generate a reduced but high-quality training database for ANN models. The resulting ANNs deliver accurate predictions of significant wave height, peak period, and wave direction, with performance comparable to the numerical model but with negligible computational cost at inference time. The results demonstrate that even ANNs trained on small MDA-derived subsets can deliver robust forecasts, making this approach suitable for near-real-time applications in operational settings. This methodology offers a scalable and computationally efficient solution for coastal wave forecasting, with implications for maritime navigation, port safety, and coastal risk management.