Motion-Compensated Spatio-Temporal Delay–Doppler Cube Detection for Robust Maritime OTFS Sensing under Nonstationary Sea Clutter

Maritime anomaly detection with Orthogonal Time-Frequency Space (OTFS) sensing commonly involves assessing individual delay-Doppler (DD) maps using set thresholds, Constant False Alarm Rate (CFAR) methods, or detectors that learn frame by frame. However, sea clutter is naturally unstable, changing with sea conditions, wind, and movement of the platform. It frequently generates brief bursts that may seem like anomalies in a single DD image, particularly when Doppler is spreading. Consequently, false positives can change, and detection accuracy becomes unreliable as conditions change. A significant drawback of many OTFS sensing systems is that they handle each DD map in isolation, failing to utilize the temporal consistency of actual anomaly signatures. This makes it hard to differentiate between lasting anomalies and temporary clutter changes in tough situations.

We introduce a spatio-temporal OTFS sensing system that detects objects using a brief series of delay-Doppler maps, instead of just one frame. The receiver creates a DD "cube" by arranging L successive OTFS frames (usually L = 3-100) and uses motion-sensitive Doppler alignment. This involves estimating the general clutter Doppler shift for each frame (for instance, by tracking the Doppler centroid or ridge) and compensating for this shift before combining the frames temporally. The aligned DD cube is then analyzed by a spatio-temporal detector, such as a streamlined 3D U-Net (a 3D convolutional encoder-decoder) or a 2D U-Net enhanced with a Convolutional Long Short-Term Memory (ConvLSTM) bottleneck or temporal attention. Training employs standard DD representations with Binary Cross-Entropy (BCE) plus Dice supervision, along with a temporal-consistency regularizer to reduce flickering detections. Potential peaks are identified using Non-Maximum Suppression (NMS) and confirmed using a track-before-detect persistence gate (for example, requiring detections in 2 of the last 3 frames along a physically reasonable drift), boosting dependability without needing manual adjustments. The suggested method, which considers both space and time, should offer small but reliable enhancements in challenging environments compared to single-image detection. These environments include situations with a poor signal-to-noise ratio, significant Doppler spread, and changes in sea conditions. In these cases, we aim for an increase in the F1-score of +0.01 to +0.03 (in situations where performance is not already at its peak) and a relative decrease of 15–25% in false positives while maintaining the same recall rate. This is mainly achieved through motion-compensated temporal fusion and persistence validation. Additionally, across different sea conditions, we seek a 10–20% relative decrease in performance variation, which would indicate greater resilience and more consistent operational performance.

The suggested system enhances OTFS maritime sensing in the face of Doppler-dispersive, nonstationary sea clutter by integrating motion-compensated temporal alignment with DD-cube detection and persistence-based confirmation. This leads to more consistent anomaly detection, avoiding the need for threshold adjustments for each specific situation.

Acknowledgment: This research was supported by the Korea Institute of Marine Science & Technology Promotion (KIMST) funded by the Ministry of Oceans and Fisheries, Korea (RS 2021-KS211502, RS-2022-KS221620).