Hybrid Panel Data Forecasting for Coastal Flood Hazard Assessment: A Spatial Data Mining Approach Using Exponential Smoothing and Neural Networks

Due to climate change–induced sea level rise, coastal flood risk is increasing significantly, creating an urgent need for improved flood risk assessment and mitigation strategies. Effective flood analysis requires both time series forecasting and hazard risk evaluation, including the determination of appropriate weights for flood-related parameters such as elevation, runoff, distance to the shoreline, and sea level. Traditional forecasting methods like exponential smoothing are limited in capturing relationships between variables, while neural networks can model non-linear interactions but are less commonly applied to long-term forecasting. To address these limitations, this study proposes a hybrid method that integrates Exponential Smoothing (ES) and Neural Networks (NN) for panel data analysis, where ES identifies trends and seasonal patterns and its output is used as additional input for NN. The NN is also employed to objectively determine parameter weights, reducing the subjectivity of conventional AHP-based approaches. The method integrates geographic and climatic variables, including wind speed, temperature, sea surface pressure, and rainfall, and is applied to coastal areas in Semarang, Demak, and Jepara, Indonesia. Results show that the hybrid model outperforms standard ES and NN methods, achieving flood forecasting errors (MAPE) between 3.5% and 8.3% and parameter weighting accuracy of 88–94%, contributing to a more holistic and reliable flood risk analysis.