Estimation of high-resolution tsunami inundation depth using deep learning models: Case study of Pangandaran, Indonesia

• An AI model that uses a low-resolution offshore tsunami height to predict high-resolution inland tsunami inundation features. • Several experiments with varying training samples and input resolutions to determine the optimal settings of the AI model. • We validate the inundation maps generated by the developed model using field survey data from historical tsunami events. Tsunami inundation maps are essential for early warning systems, guiding community evacuations to safe zones. However, producing high-resolution tsunami inundation maps is challenging due to the computational demands of numerical simulation. Thus, we propose a deep learning (DL) approach for real-time high-resolution tsunami inundation depth prediction. Our study focused on Pangandaran, Indonesia, which suffered a devastating tsunami in 2006 and remains at risk from future events in the Sunda Subduction zone. Before real-time application, the DL model requires training with low-resolution tsunami height as inputs and high-resolution inundation maps as outputs. The pre-computed tsunami training models were generated from highly diverse earthquake sources with 1044 scenarios via numerical simulations. We conducted 15 training experiments using a U-Net model with different sample sizes (1044, 800, 600, 400, and 200) and input resolutions (81, 27, and 9 arc-second). The trained U-Net using 1044 samples and 27 arc-second input resolution achieved reasonable accuracy compared to the numerical forward model as a reference, indicated by 0.357 of mean-square-error (MSE), 0.973 of intersection over union (IoU), and 0.983 of F1 score, taking less than 3 min of computational time. A retrospective forecast test for the 2006 Java tsunami demonstrated that the DL model reasonably reconstructs the reference inundation map (MSE of 0.45, IoU score of 0.97, and F1 score of 0.99) and predicts inundation heights comparable to observed values ( K number of 1.13). Our proposed U-Net model offers reliable accuracy and computational efficiency, establishing it as a critical tool for timely early warnings.