Abstract:

Submarine landslides, as a natural disaster with immense destructive potential and widespread distribution worldwide, often pose significant threats to the safety of human lives. The massive waves generated by these landslides can cause severe damage to marine structures. Therefore, rapidly predicting and assessing the size of the waves generated by underwater landslides is a crucial part of disaster prevention and mitigation efforts, essential for the development and utilization of marine resources, as well as for the safety of human lives and properties. In this study, numerical simulations of landslide-generated waves were conducted using the non-hydrostatic wave model NHWAVE (a non-hydrostatic wave model) to obtain wave data for different types of landslides. Subsequently, a CONV1D (1-dimensional convolutional neural network) was trained as the prediction model for landslide-generated waves. The model was trained using datasets for various monitoring points and different types of landslides, and evaluation metrics such as mean squared error were employed to assess the prediction performance of the convolutional neural network. The results indicate that, under the condition of using a limited amount of data, the convolutional neural network can effectively learn the patterns of landslide-generated waves. Moreover, it can predict reasonably well even for features that are not uniquely present in the dataset, demonstrating good generalization capability. Once the model is trained, inputting real-time water level data from the location of the landslide occurrence enables the neural network to predict the temporal wave profiles at downstream monitoring points in a short time. By using neural networks for prediction, it is possible to assess disasters in advance and take timely and effective response measures.

Key words: submarine landslide, convolutional neural network, temporal prediction, disaster warning