Surface water monitoring is fundamental for managing water resources and sustaining life in arid regions worldwide. Egypt is an arid country in Northern Africa, and the surface water is critical for ecosystem health, agricultural production, and human livelihoods. During dry seasons, Egyptian water bodies exhibit unique challenges for remote sensing detection due to their significant spectral differences, complex morphological patterns, and numerous small streams. However, existing water detection methods face challenges in accurately identifying water bodies with high spatial and spectral variability, especially in arid regions during dry seasons. This paper proposes an improved U-Net model with multi-scale information and attention mechanism for precise surface water mapping by multispectral Sentinel-2 satellite images. The extraction accuracy can be improved by combining convolutional layers for local feature extraction with Vision Transformer using Manhattan selfattention for global context information. Our model attains optimal performance with IoU, F1-score, recall, and precision reaching 94.26%, 97.05%, 98.18%, and 95.94%, respectively, compared to traditional machine learning methods, particularly in challenging areas with small water bodies, complex backgrounds, and eutrophic water boundaries. Our results demonstrate that integrating multi-scale information with channel and spatial attention mechanisms can effectively address the challenges of water extraction from arid environments. This advancement in remote sensing-based water extraction could enhance water resource management in arid regions globally, contributing to the monitoring and conservation of precious water resources amid increasing environmental variability.