A Dual-Branch Hybrid Deep Learning Framework with Feature-Level Fusion for Environmental Disaster Classification from UAV Imagery

Natural disasters such as wildfires, floods and building collapses are increasing risks to human life and essential infrastructure all over the world, that requires automated classification systems to provide actionable intelligence almost in real time. The Unmanned Aerial Vehicles (UAV) have been noted to be amongst the most operationally desirable platforms in disaster monitoring with a high spatial resolution, quick deployment, and access to areas hard to be accessible. Current single-branch deep learning models do not simultaneously represent large contextual scene-level features and small-scale local discriminative features, which constrains the classification of these models. This research proposed a new hybrid framework dual-branch hybrid with ResNet50 and EfficientNetB4 as a global and local feature extractor, respectively, and concatenation and fully connected classification layers as a fusing method. 5,946 UAV images of four disaster categories that include wildfires, floods, building collapses, and normal scenes were used in this experiment, Theses images are used to train and evaluate the model. The proposed hybrid model has the accuracy of 95.30% and a macro F1-score of 95.41%., which is significantly higher than standalone ResNet50 (91.40%) and EfficientNetB4 (92.70%) baselines, trained the same conditions.