TY  - EJOU
AU  - Alsolai, Hadeel 
AU  - Ahmed, Muhammad Waqas 
AU  - Alabdullah, Bayan 
AU  - Alhayan, Fatimah 
AU  - Alonazi, Mohammed 
AU  - Jalal, Ahmad 
AU  - Park, Jeongmin 

TI  - Semantic Context-Aware Multi-Scale Vision Transformer for UAV Disaster Scene Classification and Uncertainty-Aware Understanding
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - Robust scene-level classification and semantic understanding from aerial and disaster-related imagery are essential for intelligent vision systems deployed in emergency response, UAV-based monitoring, and safety-critical environments. However, existing deep learning approaches, including convolutional neural networks and Vision Transformers (ViTs), often struggle to simultaneously capture fine-grained local object characteristics and global semantic scene context, while also lacking reliable uncertainty estimation mechanisms for trustworthy decision-making. To address these limitations, this paper proposes MS-SLCA-ViT, a novel multi-scale scene–local cross-attention Vision Transformer framework for robust and uncertainty-aware image scene understanding. The proposed architecture introduces three major contributions. First, a Multi-Scale Patch Tokenizer (MSPT) extracts hierarchical semantic representations at multiple spatial resolutions and adaptively integrates them through a learnable cross-scale attention mechanism to enhance contextual feature modeling. Second, a Scene–Local Cross-Attention (SLCA) module employs dual-stream transformer processing with bidirectional attention to model interactions between global scene semantics and discriminative local object regions, improving semantic context understanding in complex environments. Third, an uncertainty-aware inference strategy based on Monte Carlo Dropout generates calibrated confidence estimates alongside predictions, enabling more reliable and interpretable decision-making for intelligent visual analytics. Extensive experiments conducted on a public disaster image dataset demonstrate that the proposed framework achieves a top-1 classification accuracy of 93.7%, outperforming state-of-the-art approaches, including ViT, Swin Transformer, DeiT, ResNet-50, and EfficientNet-B0. Furthermore, the uncertainty modeling strategy significantly improves prediction reliability and confidence calibration under challenging environmental conditions. The experimental findings demonstrate that the proposed framework provides an effective and generalizable solution for robust scene understanding, semantic context modeling, and intelligent image analysis in UAV-assisted and safety-critical computer vision applications.
KW  - UAV video scene understanding; semantic context modeling; vision transformer; multi-scale feature representation; transformer-based perception; intelligent aerial surveillance; safety-critical visual analytics

DO  - 10.32604/cmc.2026.085838