Vision-Language Model-Based System for Automated and Proactive Construction Fence Management to Enhance Severe Weather Preparedness

Abstract Construction fences can become unstable during high-wind events, particularly when attached objects such as signage, banners, or shade cloths increase the surface area exposed to wind. However, current inspection and preparedness practices often rely on manual observation, limited time and labor, and practitioners’ field experience, making it difficult to quickly assess fence stability and determine appropriate preparedness actions before high-wind events. To address this challenge, this study developed a vision-language model-based system that integrates prompt-based information extraction, computer vision-based fence condition assessment, engineering-based stability analysis, and large language model-based practical guideline generation. The system extracts key variables from user input, identifies fence panels and attached objects using instance segmentation, estimates wind-exposed surface areas, performs engineering-based stability analysis, and generates practical guidelines based on the stability analysis results. Technical evaluation showed that the instance segmentation model achieved an overall mAP50 of 0.866 in detecting fence-related objects. For practical guideline generation, the Construction Material and Incomplete and Temporary Structure 3-30E model with a temperature of 0.5 and top-p of 0.7 achieved the highest bidirectional encoder representations from transformers score F1 value of 0.853 among the tested configurations. In addition, a practitioner-based evaluation was conducted to examine the practical value of the proposed system from the perspective of potential users. The results indicated positive perceptions, particularly for practical guideline generation, system usefulness, and future adoption intention. This study contributes to automated inspection and proactive construction site management for severe weather preparedness by demonstrating how image-based condition assessment, engineering-informed stability evaluation, and natural-language decision support can be integrated into an artificial intelligence–based decision-support tool.