Deep Learning-Based Super-Resolution for Efficient Tsunami Inundation Modeling: A Case Study in Lima, Peru

Abstract

Introduction: High-resolution tsunami inundation maps are essential for coastal hazard assessment and evacuation planning; however, their generation remains computationally demanding due to the high cost of numerical simulations. This limitation restricts the rapid production of detailed inundation scenarios required for effective risk management. In this study, we propose a deep learning-based super-resolution framework to reconstruct high-resolution tsunami inundation fields from low-resolution numerical simulations.

Methods: The proposed approach employs a U-Net architecture trained on synthetic datasets generated using the TUNAMI model, focusing on coastal areas of Lima, Peru. The model learns the mapping between coarse-resolution inputs and fine-resolution inundation patterns derived from physics-based simulations. This strategy enables the approximation of high-resolution results without the need for computationally expensive numerical runs.

Results: Preliminary results indicate that the model is capable of reproducing key spatial features of tsunami propagation and inundation, while significantly reducing computational requirements compared to traditional high-resolution simulations. The approach shows particular potential in preserving spatial detail relevant for urban coastal environments.

Conclusions: These findings highlight the potential of integrating deep learning techniques into tsunami modeling workflows to enhance computational efficiency and scalability. The proposed framework can support faster generation of inundation scenarios and may serve as a complementary tool for advanced tsunami hazard assessment and early warning applications.