A physics–data hybrid framework for timely warning of landslide propagation using two-phase depth-integrated SPH model and deep learning

Landslide propagation can cause severe downstream damage, yet timely warning remains difficult because high-resolution propagation data consistent with physical laws are rarely available from field observations. To address this problem, this study proposes a physics–data hybrid framework for hazard-level warning of landslide propagation by combining two-phase depth-integrated Smoothed Particle Hydrodynamics (SPH) simulations with a Hybrid Deep Residual Convolutional Neural Network–Long Short-Term Memory (DRCNN-LSTM) model. A real debris-flow case in Hong Kong is used as the application example. The SPH model is employed to generate physically consistent propagation scenarios, from which early-stage signals, including flow height, velocity, and discharge, are extracted from the SPH simulations through an upstream monitoring window. Based on these data, a multimodal prediction task is constructed, in which short-term temporal sequences are combined with key scalar descriptors to infer the hazard level at the downstream target area. To solve this task, the proposed Hybrid DRCNN-LSTM architecture is developed with an Artificial Neural Network (ANN) branch for scalar inputs, a deep residual CNN module for temporal feature extraction, and a Long Short-Term Memory (LSTM) branch for sequential dependency learning. In this way, both static controlling information and dynamic propagation responses are jointly exploited for warning-oriented prediction. Benchmark and ablation results show that the proposed framework improves classification accuracy from 89.5 % to 94.5 %, outperforming ANN, Hybrid CNN, and Hybrid CNN-LSTM. The results further demonstrate the importance of multimodal learning and explicit temporal modelling, and confirm that early upstream monitoring information can provide effective warning of downstream hazard severity. In addition, the Hybrid DRCNN-LSTM requires less than 10−4 s per sample for prediction, indicating strong potential for rapid warning-oriented applications. This study establishes a practical route from physics-based landslide-propagation simulation to data-driven hazard-level prediction for rapid warning approaches. © 2026