Improving UAV disaster response with DenseNet and AF-RCNN: a framework for accurate emergency spot detection

Unmanned aerial vehicles (UAVs), commonly known as drones, have emerged as valuable tools in disaster response due to their ability to access remote and hard-to-reach areas. Equipped with camera sensors, UAVs facilitate real-time monitoring of disaster-stricken regions, including collapsed buildings, floods, and fires, enabling faster mitigation efforts. However, integrating deep learning (DL) models into UAVs for disaster detection introduces significant computational overhead, limiting their use in low-latency scenarios where rapid decision-making is critical. To address this challenge, this paper proposes a novel model for precise emergency spot detection that combines feature extraction from DenseNet, hyperparameter tuning using Penguin Search Optimization (PESO), and classification through an Augmented Feature Regional Convolutional Neural Network (AF-RCNN). The model introduces several innovations, including the Small Spots Feature Enrichment Extractor (SSFEE) for enhanced detection of small-scale disaster spots, the SCM Loss function for balancing feature uniqueness and commonality across varying spot sizes, and a one-to-one computation approach for optimized data sampling. Furthermore, Duck Swarm Optimization (DSOA) is employed to fine-tune the AF-RCNN parameters. The proposed model is evaluated using the AIDER dataset, and experimental results demonstrate superior performance compared to existing models, achieving faster processing speeds—up to 20 times quicker—while maintaining or surpassing accuracy. These findings highlight the potential of the proposed model for real-time, efficient emergency detection in disaster response scenarios.