Meta-Learning for Adaptive Sea Clutter Suppression: An Unsupervised Range-Doppler Domain Reconstruction Method

With the increasing demand for surveilling low-observable targets, dim target detection in complex ocean environments remains a challenging issue. Sea clutter suppression can significantly enhance target detection performance. However, the complexity and variability of the ocean environment cause the sea clutter deviate from prior assumptions, leading to suboptimal performance of empirical model-based suppression methods. Meanwhile, data-driven sea clutter suppression methods often fail to generalize across varying sea state levels. In this work, we propose an adaptive sea clutter suppression method based on meta-learning. The proposed approach takes the range-Doppler spectrum of sea clutter as a multichannel input to the autoencoder and employs unsupervised reconstruction as a proxy task to meta-train the base model. Under different environmental conditions, the model can update using few-shot samples from the unseen scenario to adapt to the current sea clutter suppression task, demonstrating strong generalization ability. The proposed method surpasses representative sea clutter suppression methods in terms of signal-to-clutter ratio improvement and detection probability. It also exhibits superior performance on the intersection over union metric. Generalization ability test results indicate that the proposed method effectively suppresses sea clutter across various sea state conditions.