A vegetation-aware deep learning model for landslide prediction with antecedent rainfall

Antecedent rainfall and vegetation both play important roles in regulating the hydrological conditions of soil, therefore influencing slope stability. However, the spatiotemporal variations of antecedent rainfall and vegetation are often overlooked in landslide prediction models. In this study, a novel deep learning model is proposed, which incorporates the effects of antecedent rainfall and vegetation to enhance the accuracy of landslide predictions. The effects of antecedent rainfall are extracted from hourly rainfall time series based on an interpretable neural network. The vegetation effects are incorporated by monitoring leaf area index from satellites and daily meteorological records from weather stations. Validation against historical landslide events demonstrates that the proposed novel model significantly outperforms existing machine learning models in terms of accuracy. The new model’s performance improves by 12·0% when 14 day antecedent rainfall is considered, confirming the critical role of antecedent rainfall in landslide prediction. Moreover, incorporating the vegetation effects further enhances the model’s performance by 4·5%. As demonstrated through a case study of the storm between 6 and 7 June 2008 in Hong Kong, the proposed deep learning model effectively reduces both false alarms and missed alarms, providing superior spatiotemporal predictive capabilities.