A Fully convolutional network for spatiotemporal feature extraction and detection of active sonar target echoes

Conventional active sonar target detection methods primarily rely on echo amplitude information, often neglecting the valuable spatiotemporal structural features (ST-SF) introduced by signal processing operations such as matched filtering and beamforming. This limitation restricts their detection performance in challenging environments. To address this issue, this paper proposes a novel method termed FCN-STFED (Fully Convolutional Network for Spatiotemporal Feature Extraction and Detection). The core of the method is an encoder-decoder based fully convolutional network (FCN), which learns, in a data-driven and end-to-end manner, a complex nonlinear mapping from time-angle energy matrix, incorporating local contextual information, to detection statistics. This enables effective exploitation of the spatial structure characteristics of target echoes. The detection threshold is adaptively determined via Monte Carlo simulation according to a preset false alarm probability. Experimental results demonstrate that the proposed FCN-STFED method achieves superior performance over the conventional two-dimensional constant false alarm rate (CFAR) detector. It yields an average improvement in detection probability of approximately 21$\%$ at the same false alarm rate when processing complex multi-highlight target models under low signal-to-clutter ratio (SCR) conditions. Visualization analyses further confirm that the FCN successfully learns discriminative structural features of targets, leading to more reliable detection under low SCR conditions. This study validates the significant potential of using deep learning to exploit inherent ST-SF in the signal processing domain, offering an effective and engineering-feasible new approach for enhancing active sonar detection performance in complex environments.