Under climate change, forest tourism sites face increased risks from extreme weather like heavy rainfall and typhoons, leading to landslides, vegetation degradation, and safety issues.Current models struggle to accurately assess these risks due to a lack of asynchronous and multi-scale temporal dynamics understanding.This paper proposes a risk assessment and early warning method using a multi-channel long short-term memory (LSTM) network and an asynchronous attention alignment mechanism.This approach enhances the perception of early nonlinear signals of extreme events.A multi-level responsive early warning model is built to achieve spatiotemporal risk mapping.Results show an area under curve (AUC) of 0.874, F1 value of 0.817, and a median early recognition time of 3.4 h, significantly outperforming existing models.The model achieves an 83.6% detection success rate within 1 h of events, with a 23.9 6.6 min response delay, improving climate risk response and decision support.