A deep-learning based high-gain method for underwater acoustic signal detection in intensity fluctuation environments

•We propose the deep-learning based separation network for underwater acoustic signal detection, which is addressed for the first time in underwater scenario by a data-driven separation method up to our knowledge.•With the help of Encoder-Decoder design, we project the input signal into a high-dimensional latent space where informative features are extracted and preserved, leading to better robustness to the random range migration compared to conventional handcraft methods.•We theoretically revise the conventional mean square error loss, which is agnostic to correlation coefficients. To address this issue, we propose a novel scale invariant signal-to-noise ratio as our loss function to better leverage the temporal coherence.