HBO-DEViT: Vision Transformer Based Attention-Guided Evolutionary Architecture for Ship-Iceberg Categorisation in Arctic Sar Images

Remote offshore regions face significant risks from drifting icebergs. Traditional instruments like Unmanned Aerial Vehicles (UAV) are impractical due to harsh weather conditions. Utilizing Synthetic Aperture Radar (SAR) imagery for iceberg monitoring is a viable solution, derived from dual-polarized Sentinel-1. This paper proposes a novel Hybrid Bayesian Optimized Differential Evolution based Vision Transformer (HBO-DEViT) model for efficiently detecting and categorising ships and icebergs in SAR images. The developed model showed promising results for iceberg and vessel detection, achieving near 99.1% accuracy on the Statoil/C-CORE dataset. This research has a significant impact on navigational safety and possesses broad applicability to address threats in the maritime domain.