Accurate estimation of glacier surface velocity using the pixel offset tracking small baseline subset (PO-SBAS) technique depends critically on the quality of synthetic aperture radar (SAR) offset observations. However, the signal-to-noise ratio (SNR) of offset measurements varies substantially across both temporal baselines and spatial locations due to glacier dynamics, seasonal surface changes, coherence loss, and topographic heterogeneity. This spatiotemporal heterogeneity complicates anomaly detection, as anomalies are defined relative to normal displacement behavior that itself evolves over time. Particularly under low SNR conditions, true anomalies are difficult to distinguish from noise and temporally varying glacier motion, limiting the effectiveness of traditional fixed threshold and statistical filtering methods, which typically assume that noise follows a Gaussian distribution. To overcome this challenge, phase-space reconstruction is employed, embedding each pixel's offset time series from different baselines into a high-dimensional phase space. This transformation reveals the underlying temporal structure of glacier motion while preserving nonlinear dynamics. Anomalies are then identified via alpha shape analysis, which extracts the dense core trajectory and adaptively flags outliers as deviations beyond a compact spatial envelope. We evaluate the method using Sentinel-1A data from the Gyala Peri region, spanning multiple temporal baselines. Results show that even under varying baseline lengths and complex noise conditions, the core trajectory in phase space remains stable. The alpha shape method effectively captures this structure, enabling robust and adaptive outlier removal. Compared with the traditional time-series filters, our approach better handles spatiotemporal SNR variability and preserves physically consistent offsets signals, ultimately improving glacier velocity estimation under complex observational conditions.