Retrieval and classification of disaster scenes under complex environmental conditions remain challenging tasks for UAV-based emergency response systems due to large-scale variations, cluttered backgrounds, and heterogeneous damage patterns. To address these challenges, this paper proposes a hybrid late-fusion framework that integrates global and local Convolutional Neural Network (CNN) models for UAV-based disaster image classification. The global CNN is designed to capture scene-level contextual information from entire UAV images, while the local CNN focuses on extracting fine-grained spatial details related to localized damage patterns. The outputs of both models are fused at the decision level to enhance classification robustness. Experiments conducted on a real-world UAV dataset demonstrate that the proposed hybrid framework achieves an F1-score of 0.85, outperforming individual CNN models in terms of balanced precision and recall. The results confirm that integrating global contextual features with local structural details provides a robust and effective solution for UAV-based disaster image classification, particularly for post-disaster analysis and decision-support systems.
Samir et al. (2026) studied this question.