When the Stakes Are High: Trustworthy Machine Learning for Ocean and Coastal Hazard Forecasting

The areas where coasts touch the ocean attract humanity, jobs, infrastructure and pressure from waves, storm surge, flooding and eroding shores with rising seas, stronger storms and altered weather patterns. Physical simulation (physics-based) numerical wave and surge models have been employed in forecasting such phenomena for many years. Examples are SWAN, WAVEWATCH III and XBeach. These are based on fundamental physical principles and offer good accuracy but are costly in terms of compute time. Accordingly, they do not typically produce daily predictions or finer resolutions sufficient for day-to-day operations and risk management. Nevertheless, data-driven and machine learning methods have emerged in the last few years and rapidly permeated wave and wind forecasting, shoreline mapping, analysis of historical shoreline change, bathymetry measurement and mapping coastal hazards. Their predictions match or beat traditional techniques at multiple scales, yet they generally come with the drawbacks of being opaque, of rarely linking forecasts to physical cause and rarely presenting realistic estimates of uncertainty.This Research Topic addresses what it takes for us to trust coastal and ocean machine learning when stakes are high: our lives, livelihoods and the natural environment. We want to shift our focus away from "leaderboard accuracy" and towards interpretable, physically meaningful, well-calibrated machine learning that delivers genuine value to those managing coasts. We invite papers that addressexplainability (e. g. , through SHAP, other feature attribution techniques or by incorporating physical understanding in hybrid models), meticulous validation (e. g. , chronological train-test splits, generalization across sites, quantifying uncertainties) and meaningful application contexts (e. g. , early warning systems, erosion and flood management strategies, coastal engineering and design processes). Placing ML researchers next to coastal scientists and managers may illuminate what modeling techniques actually contribute to better predictions and which only succeed on a metrics scorecard.Submissions concerning significant wave height and wind forecasting, wave runup and overtopping prediction, satellite-based shoreline mapping and change, bathymetry mapping and analysis, coastal vulnerability and hazard assessment, prediction of storms and flooding, prediction of transport of sediments or pollutants e.g., microplastics in any marine, estuarine or shelf coastal setting are welcome. Studies aiming to improve interpretability and explainability of AI models, the use of physics-informed and hybrid modeling approaches, uncertainty quantification in modeling or observation interpretation, benchmarking and reproducibility of results and the transferability and generalization of models to different regions or conditions are of particular interest. Submissions will include Original Research, Methods, Review and Mini Review articles, as well as Data and Code Descriptors to enhance the reproducibility of coastal-ocean machine learning science. Topic editors: Ahmet Durap, Zekâi Şen, Iskender Sahin, Erdal Kesgin, Eyyup Ensar Başakın Source: https://www.frontiersin.org/research-topics/82207/when-the-stakes-are-high-trustworthy-machine-learning-for-ocean-and-coastal-hazard-forecasting