Key points are not available for this paper at this time.
In the face of increasing global tourism uncertainties, timely and accurate risk identification is crucial for visitor safety and destination resilience. While image-based visual cognition is promising, traditional deep learning models lack semantic reasoning, and generalist Multimodal Large Language Models often lack the domain-specific precision required for tourism safety. To bridge these gaps, we propose the Tourism Activity Risk Identification Model (TARIM). By establishing a domain-knowledge-driven fine-tuning paradigm, this framework transforms unstructured visual data into structured decision-support information. Experimental results demonstrate that TARIM achieves a 96.83% risk recall rate and a 0.81 Macro-F1 score, significantly outperforming existing baselines. This study proposes a domain-oriented multimodal framework for real-time safety monitoring via ubiquitous sensing, effectively identifying environmental hazards, facility risks, and behavioural anomalies. By integrating image cognition with visual cue reasoning, TARIM overcomes the limitations of rule-based detection. This research provides a robust, lightweight methodological framework for enhancing emergency response timeliness, thereby contributing to the high resilience and sustainable development of tourism destinations.
Liu et al. (Tue,) studied this question.