IoT-based intelligent emergency logistics management using DETR-driven visual perception and deep Q-learning

Emergency logistics systems face significant challenges in disaster scenarios due to damaged infrastructure, unpredictable conditions and the need for rapid adaptive decision-making. Traditional rule-based models lack the situational awareness and real-time intelligence required for efficient resource coordination under such constraints. This paper proposes an integrated framework that combines DETR-driven visual perception with Deep Q-Learning to achieve end-to-end intelligent emergency logistics management. High-resolution satellite imagery is processed using a fine-tuned Detection Transformer (DETR), which directly predicts object classes and bounding boxes without relying on region proposals. Extracted objects are passed through ResNet-50 to generate 2048-dimensional semantic feature vectors, preserving contextual and structural information critical for decision-making. These features are concatenated with dynamic environmental parameters and used as input states in a Deep Q-Network (DQN) which models the decision environment as a Markov Decision Process. The DQN learns optimal policies for logistics actions such as supply rerouting and aerial deployment through exploration, exploitation and temporal-difference learning. A prioritized experience replay buffer and target network are used to stabilize training, while reward functions drive policy convergence towards efficient emergency responses. The entire pipeline is designed for real-time responsiveness and continuous environmental adaptation. The Internet of Things system collects real time information and data using environmental sensors, which feed into the perception and decision-making elements. Experimental evaluation using the xView dataset demonstrates detection accuracy of 88.56%, feature representation accuracy of 90.14% and action recommendation accuracy of 86.78%, with low decision latency. Compared to state-of-the-art baselines the system achieves higher operational performance, making it suitable for real-world deployment. This work has the potential to significantly enhance the agility and intelligence of disaster logistics operations in both government and industrial emergency response systems.