An innovative deep vectorised scaling Swin connectivity U-Net attention network for disaster detection in satellite imagery

The growing frequency and intensity of natural calamities, including floods, wildfires, and hurricanes, demand accurate and robust disaster detection systems built on satellite images. The proposed paper will solve these issues by a new Deep Vectorised Scaling Swin Connectivity U-Net Attention Network (DVSCUAN) framework combined with an Artificial Circulation System Algorithm (ACSA) to achieve adaptive optimisation. The proposed architecture uniquely integrates vectorised feature scaling, Swin Transformer–based connectivity, and attention-guided U-Net decoding to enhance multi-scale feature representation and spatial coherence in disaster segmentation. Also, two-level diffusion-based denoising mechanism is also used to enhance resistance to noise and degradation of resolution. DVSCUAN generalisability is justified in various disaster problems, such as flood, wildfire, and hurricane damage detection, based on a wide range of benchmark datasets, including FloodNet, Wildfire Prediction Dataset, xBD, and SIHD. Extensive experimental analyses show that DVSCUAN always outperforms state-of-the-art models with 98.9% accuracy, 98.6% F1-score, 97.4% IoU, and 97.8% Dice score, and it is also computationally efficient.