<p>Landslide-induced tsunami hazards pose a significant threat to marine resource development and the safety of coastal populations. Accurate forecasting of landslide-generated tsunami waves using effective methods and models is therefore of critical importance. Previous studies have applied one-dimensional convolutional neural networks (CNN1D) and multilayer perceptron (MLP), for tsunami prediction. However, tsunami data are often difficult and expensive to obtain, and training models using large volumes of data is frequently impractical under real-world application conditions. To address this challenge, this study proposes a slime mold algorithm-optimized extreme learning machine (SMA-ELM). and evaluates its performance in small-sample end-to-end prediction tasks using K-fold cross-validation, comparing it with CNN1D and MLP models employed in prior research as well as mainstream machine learning models including random forest (RF), extreme gradient boosting (XGB), and kernel extreme learning machine (KELM). Experimental results indicate that the average RMSE (Root Mean Square Error) of SMA-ELM under small-sample conditions is 0.001, significantly lower than that of CNN1D (0.0039), MLP (0.04), KELM (0.019), and RF/XGB (0.043). This demonstrates that the proposed model achieves superior prediction accuracy and stability. Moreover, even when label characteristics and prediction samples are altered, the SMA-ELM model maintains accurate forecasting capabilities at target locations, highlighting its strong robustness and reliability. Additionally, the ELM model features a simple, efficient, and reliable architecture, enabling it to deliver more valuable early warning time for tsunami events and demonstrating significant potential for engineering applications.</p>