Underwater Object Detection encounters persistent challenges caused by spectral attenuation, forward and backscatter, and severe loss of high-frequency detail. These degradations suppress edges and fine structures that govern tiny-object visibility. Such effects also disturb spatial sampling by creating unstable local geometry in low signal-to-noise regions. Most recent methods strengthen semantic context and adopt generic deformable sampling, yet the alignment between frequency restoration and transmission-guided feature resampling is still underexplored. This paper introduces FTAS-YOLO, a unified Frequency Transmission Adaptive Sampling framework. The proposed AquaFreqFuse performs low–high frequency decomposition and channel attenuation modeling, whereas HydroOffsetNet conducts transmission-aware Gaussian sampling during feature upscaling. This design restores attenuated structure and stabilizes spatial reconstruction under heterogeneous water conditions. Experiments across multiple underwater benchmarks show consistent improvement over existing detectors in both accuracy and robustness. The method delivers strong small-object recovery in clear and turbid environments and maintains practical runtime suitable for real-time deployment on underwater platforms.