On 29 July 2025, an Mw 8.8 megathrust earthquake occurred off the eastern coast of the Kamchatka Peninsula. The earthquake generated a large tsunami that reached Japan approximately 2 h later, with a maximum observed height of ~1.3 m. Offshore observation networks have been deployed in Japan: S-net, DONET, and N-net. These networks are equipped with bottom pressure sensors that record water pressure in real time, enabling tsunami detection before waves reach the coast. Offshore observations can be adopted to predict coastal tsunamis based on data assimilation approach, which has been successfully validated using linear propagation models. Although linear models can accurately predict tsunami arrival, they are less capable of forecasting the tsunami decay process (i.e., following waves). The 2025 Kamchatka earthquake tsunami was the largest event since the completion of S-net, providing an opportunity to extend the data assimilation framework to a nonlinear formulation. We applied a nonlinear propagation model that accounts for advection and bottom friction, using S-net observations as the inputs for data assimilation. The objective was to predict tsunami arrival and decay process. The Tohoku region of Japan was selected as the study area, and the tsunami was predicted up to 24 h after the earthquake. Results show that data assimilation approach enables accurate predictions of tsunami heights at least 30 min before its arrival. Linear and nonlinear models perform comparably in predicting tsunami arrival. However, for the following wave prediction, the linear model tends to diverge due to numerical instabilities, whereas the nonlinear model more reliably captures decay process. The nonlinear model achieved prediction accuracy of 60–80% across eight tide gauges, with an average RMSE between observed and forecasted waveforms of