Low-Latency Marine-Based OTFS Echo Parameter Estimation Enabled by AI

We propose an end-to-end pipeline for Orthogonal Time-Frequency Space (OTFS) sensing that integrates deterministic signal processing with a Machine-Learning (ML) inference stage. The pipeline first generates a complex delay-Doppler grid via standard Symplectic Fast Fourier Transform (SFFT)-based OTFS reception. We then employ an 'oracle' Ground-Truth (GT) association process to deterministically label signal peaks, extracting their complex gain (α) and absolute indices (m,n) to deduce physical targets (range, radial velocity). These oracle-aligned labels are used to train a Random-Forest (RF) classifier. The RF model learns to map normalized 33×33 complex patches, centered on signal peaks, to their corresponding target parameters. On an 80/20 split of 10,000 samples, the classifier achieved a 0.966 accuracy, 0.965 macro-F1 score, and 0.998 macro Receiver Operating Characteristic-Area Under the Curve (ROC-AUC). Notably, when tested on held-out scenes, the model's derived range and velocity predictions achieved 100% coincidence with the GT, while amplitude and phase corresponded in 89% of instances. This hybrid oracle-and-ML approach demonstrates a highly effective and robust method for precise target extraction in OTFS-based sensing systems.