Unsupervised Domain Adaptation (UDA) poses a major challenge in computer vision, where models trained on a source domain often fail on target domains due to distributional shifts. We propose Domain Adaptation using Vision Transformers with Instance-Level Discrimination and Statistical Alignment (DA-VIDSA)—a novel Vision Transformer (ViT-16)-based framework that combines classification loss with global (MMD, CORAL), class-conditional (LMMD), and instance-level discriminative alignment to learn robust, domain-invariant features. Evaluated on the Fruit-360 dataset across four cross-domain tasks with varying class counts and bidirectional shifts (FO \(\leftrightarrow \) FP), DA-VIDSA consistently outperforms existing methods, achieving up to 94.11% accuracy in the most challenging settings. Our results demonstrate the effectiveness of integrating multi-level alignment with ViTs for scalable, label-efficient domain adaptation.