<p>Underwater image enhancement plays a crucial role in diverse fields such as marine biology, underwater archaeology, and autonomous underwater vehicle operations. Despite its importance, existing enhancement methods often struggle with challenges posed by light attenuation and scattering, resulting in blurry images with distorted colors. To address these issues, this study proposes RASS-U-Net, a novel deep learning framework that integrates four specialized modules designed for underwater environments: a ResNet-based underwater-adapted encoder, Atrous Spatial Pyramid Pooling (ASPP) for multi-scale context aggregation, Selective Kernel Convolutions (SKC) for adaptive feature fusion, and a Spatial Attention Mechanism (SAM) for guided restoration. Inspired by but redesigned from conventional ResNet, ASPP, SKC, and SAM architectures, each component specifically targets underwater image degradation. The encoder employs residual learning with spectral filtering and fine-tuning to enhance low-level feature extraction while preserving color fidelity. The ASPP module captures contextual information at multiple degradation scales using customized atrous convolutions. In the decoder, SKC adaptively selects receptive fields to handle spatially varying distortions, while SAM focuses restoration on regions affected by turbidity and contrast loss. Extensive experiments conducted on the UIEB and EUVP datasets demonstrate that RASS-U-Net significantly improves image clarity, color fidelity, and detail preservation. Evaluation metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Underwater Color Image Quality Evaluation (UCIQE) confirm that the proposed framework surpasses current state-of-the-art methods in delivering enhanced underwater images with higher quality and more accurate colors. This work provides a robust and adaptable solution for underwater image enhancement, facilitating more precise and reliable imaging for various scientific and practical applications.</p>