Underwater images are often plagued by low brightness, contrast, and blurred vision due to sediment and light scattering, hindering effective feature extraction. This paper proposes SGNeR, a novel underwater image enhancement model leveraging semantic segmentation and implicit neural representation. SGNeR combines real and augmented scenes to retain more image details, utilizing semantic segmentation as supervision to improve image quality and segmentation accuracy. To reduce dependence on paired training data, a triple-closed-loop constraint module enables self-supervised learning. Thorough experiments demonstrate that SGNeR outperforms state-of-the-art methods, achieving superior visual and quantitative results on various benchmarks. Our approach offers a robust, data-driven solution for underwater image enhancement, with promising applications in marine exploration and research. The dataset we used and the example code are publicly available at this link : https://github.com/ssssxw/SGNeR