准确 delineate 内陆水体并监测地表水动态变化对于水文研究和气候适应具有重要意义。表面水与海洋地形（SWOT）卫星显著提升了全球地表水观测能力。然而，在复杂的内陆环境中，SWOT 数据常受条带噪声和质量控制（QC）标记失效的影响，易导致水体提取错误或遗漏狭窄水体。为解决上述问题，我们提出了 SDNet，一种基于 Transformer 的多尺度框架。该设计在抑制高频条带噪声的同时保留细尺度水文边界，显著提高了水体提取的可靠性。不同水体的实验结果表明：（1）SDNet 在各类尺度水体上均实现了高精度。相较于基于 QC 的分类方法，大型水体的水体交集率（WIR）和背景交集率（BIR）分别提升了 23.68% 和 45.96%；中等尺度水体的 WIR 进一步提升 2.83%，小型水体的 WIR 提升达 3.57 倍。（2）利用 ICESat-2 高程数据进行交叉验证显示，SWOT 高程误差与跨轨距离呈正相关。在 100 m 分辨率下，中位误差范围为 0.117 m 至 0.181 m，在 250 m 分辨率下为 0.111 m 至 0.170 m，平均绝对误差保持在分米级以下。（3）季节性水文分析揭示了不同类型水体对水位变化响应模式的显著差异：自然湖泊主要受气候过程驱动，而调控型水库则表现出由人类调控主导的多峰动态特征。

Accurately delineating inland water bodies and monitoring surface water dynamics are crucial for hydrological research and climate adaptation. Surface water and ocean topography (SWOT) satellites has significantly improved global surface water observation capabilities. However, in complex inland environments, SWOT data are often affected by stripe noise and quality control (QC) marker failures, which can easily lead to water body extraction errors or the omission of narrow water bodies. To address these issues, we developed SDNet, a transformer-based multi-scale framework. This design suppresses high-frequency stripe noise while preserving fine-scale hydrological boundaries, thus significantly improving the reliability of water body extraction. Experimental results for different water bodies show that: (1) SDNet achieved high accuracy across water bodies of different scales. Compared to the QC-based classification, the Water Body Intersection Rate (WIR) and Background Intersection Rate (BIR) of our method increased by 23.68% and 45.96%, respectively, for large water bodies. WIR further increased by 2.83% for medium-scale water bodies and by a factor of 3.57 for small water bodies. (2) Cross-validation using ICESat-2 altimetry data showed that the SWOT altimetry error was positively correlated with cross-track distance. The median errors ranged from 0.117 m to 0.181 m at 100 m resolution and 0.111 m to 0.170 m at 250 m resolution, with the mean absolute error remaining in the sub-meter range. (3) Seasonal hydrological analysis revealed distinct response patterns across different water body types to water level changes. Natural lakes are mainly driven by climate processes, while controlled reservoirs exhibit multi-peak dynamic characteristics dominated by human regulation. These findings provide a scalable solution for multi-scale water body monitoring, contributing valuable support to hydrological research, flood risk assessment, and climate adaptation strategies.