Underwater integrated sensing and communication (UISAC) systems employ a unified signal waveform to facilitate both data exchange and sense. However, communication-centric UISAC systems face challenges due to the stochastic nature of communication signals used for sensing, leading to diminished sensing performance. Meanwhile, sonar systems, in comparison to radar systems, operate in a complex noise environment with time-varying and non-Gaussian properties, further deteriorating sensing capabilities. To address these challenges, we propose an adaptive algorithm using orthogonal time-frequency space (OTFS) signals for underwater target sensing. Our approach incorporates an online Gaussian-mixture model (GMM) to track time-varying background noise and utilizes a weighted Mahalanobis distance as a statistical measure to differentiate between reflected OTFS signals and ambient noise rather than relying solely on waveform. Comparative evaluations with existing energy detectors in UISAC system show that our strategy improves sensing accuracy, enhancing performance even in the presence of complex oceanic noise.