Fractal and Machine Learning-Based Analysis of Shoreline Change in Storm-Affected Coastal Environments

Understanding and forecasting coastline retreat under storm forcing is critical for coastal resilience planning. This study presents a multi-site, data-driven framework integrating fractal shoreline metrics, ERA5-derived storm parameters, and machine learning models to predict monthly shoreline retreat across three morphologically distinct Turkish coasts (Istanbul Karaburun, Karasu, and Sinop Boyabat). Satellite imagery (Sentinel-2, Landsat-8) was processed to extract historical shorelines, followed by computation of box-counting (D_box) and boundary-method (D_bm) fractal dimensions. Storm indicators—including storm index, high-wind hours, significant wave height exceedance durations, and wave energy metrics—were derived from ERA5 reanalysis. To address data sparsity, a block-bootstrap data augmentation strategy generated 300 synthetic years, yielding a final dataset of 10,815 monthly observations.

Five machine learning models were trained and evaluated: Random Forest, GRU, MLP, Ridge, and SVR. Results demonstrate that the Random Forest model achieved the highest performance (R² = 0.9866, MAE = 0.0113 m/month), followed by the GRU model (R² = 0.9552). Feature importance analysis revealed that shoreline-specific characteristics, fractal metrics, and storm intensity indicators are the dominant predictors. Cross-site generalization tests show that a single unified model can effectively predict retreat patterns across morphologically diverse coastlines.

The findings highlight: (1) the strong predictive capacity of storm indices and fractal shoreline complexity; (2) the suitability of tree-based and recurrent neural models for coastal change prediction; and (3) the viability of bootstrap-based synthetic augmentation for long-term coastal datasets. This framework provides a scalable and transferable methodology for early-warning systems and climate-adaptation strategies. Future work will explore scenario-based forecasting under extreme storm thresholds and SHAP-based interpretability to enhance model transparency.