MCDNet: A New Multimodal Cross-Domain Network for Classification of Challenging Scene Images

Image classification of complex scenes remains a significant challenge in computer vision due to high intra-class variability, strong inter-class similarity and domain shifts between training and testing data. Existing methods often experience substantial performance degradation when applied to images drawn from distributions that differ from the training domain. In this work, we introduce MCDNet, a Multimodal Cross-Domain Network designed to address these challenges through three key innovations: (1) a dual-stream feature extraction backbone that integrates Vision Transformers (ViT) and Convolutional Neural Networks (CNNs) to capture both global semantic context and local discriminative features; (2) a Cross-Domain Adaptive Fusion Module (CDAFM) that aligns multimodal feature representations across domains using supervised contrastive learning and (3) a Hierarchical Graph Attention Network (HGAN) that models spatial relationships among scene components to enhance structural reasoning. Extensive experiments conducted on four benchmark datasets– UC Merced Land-Use, AID and NWPU-RESISC45 — as well as a newly constructed cross-domain dataset, demonstrate that MCDNet achieves competitive performance compared with recent approaches. Notably, MCDNet attains state-of-the-art accuracy on NWPU-RESISC45 with a [Formula: see text]1.69% improvement and delivers substantial gains in cross-domain generalisation on CDS ([Formula: see text]10.61% over the strongest baseline). Ablation studies verify the individual contribution of each module, while visualisation results provide interpretable insights into the model’s decision-making process.