Low probability of detection (LPD) underwater acoustic communications are an essential requirement for command and control of autonomous underwater vehicles (AUV) or submarines performing covert missions, avoiding their detection while communicating. Based on low power signals, these covert communications may also extend the autonomy of battery-operated AUVs, and contribute to reducing the impacts of the environmental noise level on marine life. The present thesis aims to develop LPD communications based on modeled and real data from three shallow water experiments. Thus, a superimposed training method has been proposed. A bitstream is created superimposing a long probe to the message before transmission. Computationally simple, the algorithm explores temporal diversity to increase the processing gain and uses a Wiener filter for equalization. Experimental results presented bit-error rates (BER) < 10−2 for signal-tonoise ratios (SNR) < −8 dB. To understand the effects of coastal upwelling phenomena over low SNR communications, a study compares the acoustic propagation for different sound speed profiles using a propagation model and analyzes data from the BioCom’19 experiment, performed off Cabo Frio Island (Brazil). Temporal and spatial coherence of low power signals propagating in this harsh environment are estimated, and both a criterion for multichannel combining and a double Wiener filter to improve equalization are presented. Passive time-reversal (pTR) techniques have been widely employed for communications. To address the pTR channel mismatch due to the environmental variability between the probe and the data transmissions, this work proposes a superimposed training pTR (STpTR) approach for single and multichannel systems. Despite the high noise levels, varying from -5 to +6 dB, the STpTR combined with a Wiener filter achieved BER < 10−2, for bit rates up to 220 bps. To improve covert communications for AUVs, this work also presents a study about vector sensor multichannel combining. Using the STpTR approach, results from an experiment on the coast of Algarve/Portugal indicate that combining the pressure and particle velocity channels of a vector sensor may provide an average SNR and mean squared-error gain of up to 9.4 and 3.1 dB, respectively, compared to the pressure channel. Therefore, a better understanding of the environment combined with the superimposed training pTR using a vector sensor may improve the LPD communication system’s performance and robustness while keeping covertness.