Evaluating seismic landslide risks using AI models: TCN, GRU and random forest techniques – a case study

Purpose This study aims to analyze and compare the efficacy of three sophisticated artificial intelligence (AI) models – temporal convolutional network (TCN), gated recurrent unit (GRU) and random forest (RF) – in evaluating seismic landslide risk in Sichuan Province, China, an area particularly susceptible to earthquake-triggered landslides. This study aims to determine the most efficient AI methodology for precise risk prediction and to produce actionable insights for mitigating landslide dangers. Design/methodology/approach This research uses a seismic landslide inventory together with influencing elements, such as geological, topographical and seismic parameters, to train and validate the AI models. The Gini index (GI) of the RF model classified landslide risk factors according to their importance. Performance was assessed with criteria such as accuracy (ACC), recall and area under the curve (AUC). A GIS-based risk distribution map has been developed to display and assess regional vulnerability. Findings The findings reveal that all three AI models – TCN, GRU and RF – exhibited exceptional performance in seismic landslide risk assessment, with ACC and AUC values exceeding 0.75 and 0.80, respectively. The TCN model demonstrated the highest accuracy (ACC = 0.781) and robustness (AUC = 0.851), positioning it as the optimal choice for seismic landslide prediction. The RF model additionally enabled the categorization of landslide risk variables based on variations in the GI. These findings were validated using five-fold cross-validation for RF and repeated randomized validation trials for TCN and GRU. All validation frameworks maintained the model performance hierarchy (TCN > GRU > RF), with statistical testing confirming TCN’s higher ACC and AUC. The GIS-based seismic landslide risk map for Sichuan Province provides critical information for local authorities and planners, improving disaster preparedness and landslide mitigation strategies. Originality/value This study compares advanced AI techniques (TCN, GRU, RF) in seismic landslide risk assessment in Sichuan Province’s hazardous landscape. The TCN model outperforms others, enhancing AI-based disaster prediction. The GI classification and GIS-based landslide risk map provide practical tools for disaster mitigation and planning, reducing earthquake-induced hazards in high-risk regions. This study introduces the novel application of TCN for modeling earthquake-induced landslides, leveraging its strength in capturing temporal dependencies from seismic triggers. By combining this capability with rigorous statistical validation – including cross-validation, confidence intervals and paired t-tests – the study establishes a reliable and generalizable framework for AI-based EQIL assessment.