Integrating Deep Learning and Vision-Language Models for Unmanned Aerial Vehicle (UAV)-Based Disaster Response: A Survey and Experimental Study

In the last five years, wars and worsening climate events have dramatically increased the necessity for Unmanned Aerial Vehicles (UAVs) using advanced Artificial Intelligence (AI) systems.The degree of usefulness that UAVs provide will depend on the onboard AI that can transform unprocessed imagery into actionable intelligence.This study reviews the developments in AI-based UAVs for identifying hazards and managing disasters from 2019 to 2025.The progression of deep learning architectures, such as YOLO to modern transformer architectures and Vision-Language Models (VLMs), has improved significantly recently.This study aims to propose a VLM-based pipeline for real-time hazard detection, reasoning, and situational reporting in a structured format.Experiments to evaluate detection performance in complex aerial scenarios using the VisDrone2019 dataset showed that the YOLOv11s lightweight architecture was superior to YOLOv11n (mAP at 50 = 0.51 vs mAP at 50 = 0.44).This study examines the role of multimodal sensing and reasoning to increase the quality of this system in complex environments.The challenges identified include real-time processing on edge devices, multimodal data fusion, and minimizing hallucinations in VLMs.The study ties together survey results with promising research directions for the future generation of UAV-based autonomous disaster response.