提示词驱动下MLLM与YOLO的城市内涝车辆淹没识别性能对比OA

To address the technical selection requirements for vehicle submersion level identification in urban waterlogging scenarios,this study constructs a comparative evaluation framework for multimodal large language models and YOLO series algorithms.Based on a sample of 1 000 vehicle images covering five submersion levels,the recognition performance of Qwen_VL_Max and GPT-4o under zero-shot and few-shot prompting modes is systematically evaluated,with YOLOv8s,YOLOv11s,and YOLO26s serving as supervised learning baseline models for comparative analysis.The results show that:① The YOLO series algorithms generally outperform the multimodal large models,with YOLOv11s demonstrating better balance and stability in level recognition,while YOLO26s exhibits advantages in standard detection metrics.② Few-shot prompting significantly improves the recognition accuracy of Qwen_VL_Max by 22.10 percentage points compared to the zero-shot mode,yet multimodal large models still lack fine-grained discriminative capability for medium-to-high submersion levels.③ The YOLO series algorithms are more suitable for large-scale,high-precision standardized recognition tasks,whereas multimodal large models have certain advantages in small-sample,rapid-deployment scenarios.The findings provide a reference for technology selection in vehicle submersion recognition for smart flood control applications.