Effective and accurate extraction of water bodies from high-resolution remote sensing imagery is important for environmental monitoring, disaster management, and resource planning. This study introduces a Vision Transformer based method utilizing the SegFormer architecture to segment water bodies from multispectral imagery. This approach is supported by Landsat-8 data and Gaofen Image Dataset water body maps for ground-truth labelling. We developed and evaluated three variants of ViT—ViT-Small, ViT-Base, and ViT-Large which incorporate a patch embedding layer, Multi-Head Self-Attention, and an MLP-based decoder to effectively capture global spatial dependencies. We compared their performances against traditional methods as well as CNN-based architectures. The results show that ViT-Large achieves up to 97.1% accuracy, IoU of 89.8%, and F1 Score of 92.7%, significantly surpassing baseline methods. While ViT-S presents a more lightweight alternative (accuracy 95.6%, inference time of 120 ms), the increased computational cost of ViT-L (200 ms) demonstrates a tradeoff between accuracy and efficiency. These results support the benefits of transformers in modelling complex spatial relationships within high-resolution imagery, thereby enhancing the capabilities of water body extraction. We hope this research will enhance the advancement of knowledge within this field and motivate the community to create more effective and innovative models.