Surface and underwater (S/U) acoustic targets recognition is an important application of passive sonar. It is difficult to distinguish them due to the mixture of underwater target radiation noise and marine environmental noise. In previous studies, although using a single hydrophone was able to identify S/U acoustic targets, there were still a few hydrophones that had poor accuracy. In this paper, S/U acoustic targets recognition using two hydrophones based on Gradient Boosting Decision Tree is proposed, and it is first found out as high as 100% accuracy could be achieved with the implementation of SACLANT 1993 data. The real experimental data are always rare and insufficient. The big training dataset is generated using environmental information by acoustic model named KRAKEN. Simulation and experimental data used in the model are heterogeneous, and the differences between these two kinds of data are assimilated by using vertical linear array feature extraction method. The model realizes the recognition of S/U acoustic targets based on channel information besides source spectrum information. By using the combination of two hydrophones, the surface and underwater targets recognition accuracy reached 1 and 0.9384, while they are only 0.4715 and 0.5620 using a single hydrophone, respectively.

Sound source recognition is a very important application of passive sonar. How to distinguish between surface and underwater acoustic sources has always been a challenge. Due to the mixing of underwater target radiated noise and marine environmental noise, especially in shallow water environments where multipath effects exist, it is difficult to distinguish them. To solve the surface and underwater acoustic source recognition problem, this paper proposes a multi-channel joint detection method based on machine learning. First, the simulation data are generated using the normal model KRAKEN setting in the same environment as the SACLANT 1993 experiment, which uses a vertical linear array of 48 hydrophones. Secondly, the GBDT classifier and LightGBM classifier are trained separately, and then the model is evaluated using precision, recall, F1, and accuracy. Finally, four ML models (kNN, random subspace kNN, GBDT, and LightGBM) are used to analyze all 48 channels of hydrophone data. For each model, two kinds of feature extraction methods (module features, real and imaginary features) are applied. Generally, the results show that both GBDT and LightGBM models have better performance than both kNN and random subspace kNN ones. For both GBDT and LightGBM models, the results using module features have better performance than using real and imaginary features.

Underwater acoustic target localization (UATL) is challenging but has achieved some success with limitations. For example, matched field processing (MFP) is sensitive to environmental noise and inefficient in processing large-scale data, making real-time with accurate performance difficult. This paper presents a Bayesian optimization-tuned machine learning approach for UATL and conducts comparative studies with MFP and other parameter tuning methods. The environment used is from the seatrial conducted on October 26, 1993, in the shallow sea area north of Elba Island. First, the simulated training data is generated by the KRAKEN propagation code on grids of ranges and depths. Second, MFP and two machine learning methods (k-nearest neighbor, support vector regression) with distinct hyperparameter optimization approaches are employed for localization. The results show that the machine learning approaches achieve higher localization accuracy than MFP, identifying the underwater target located at a 5.6 km range (error < 0.1 km) and 79 m depth (error < 0.5 m), while Bayesian optimization proves more efficient than alternative tuning methods.

Direction-of-arrival (DOA) estimation is fundamental in array signal processing, yet classical algorithms suffer from significant performance degradation under low signal-to-noise ratio (SNR) conditions and require computationally intensive eigenvalue decomposition. This study presents a systematic comparative analysis of two backbone networks, a convolutional neural network (CNN) and long short-term memory (LSTM) for DOA estimation, addressing two critical research gaps: the lack of a mechanistic understanding of architecture-dependent performance under varying conditions and insufficient validation using real measured data. Both networks are trained using cross-spectral density matrices (CSDMs) from simulated uniform linear array (ULA) signals. Under baseline conditions (1° classification interval), both CNN and LSTM methods reach an accuracy (ACC) above 98%, in which the error is ±1° for CNN and ±2° for LSTM, only existing in the end-fire direction. Key findings reveal that LSTM maintains above 90% accuracy down to −20 dB SNR, demonstrating superior noise robustness, whereas CNN exhibits better angular resolution. Four performance boundaries are identified: optimal performance is achieved at half-wavelength element spacing; SNR crossover occurs at −20 dB below which accuracy drops sharply; the snapshot threshold of 32 marks the transition from snapshot-deficient to snapshot-sufficient conditions; the array size of 8 is the turning point for the performance variation rate. Comparative analysis against traditional methods demonstrates that deep learning approaches achieve superior resolution ability, batch processing efficiency, and noise robustness. Critically, models trained exclusively on single-target simulated data successfully generalize to multi-target experimental data from the Shallow Water Array Performance (SWAP) program, recovering primary target trajectories without domain adaptation. These results provide concrete engineering guidelines for architecture selection and validate the sim-to-real generalization capability of CSDM-based deep learning approaches in underwater acoustic environments.

Passive acoustic localization in complex shallow waters requires algorithms tailored to specific operational constraints. This paper investigates the adaptability, computational efficiency, and statistical performance boundaries of five matched-field processing (MFP) methods—Bartlett, Minimum Variance Distortionless Response (MVDR), Multiple Signal Classification (MUSIC), Reduced Covariance Matrix (RCM), and Rank and Trace Minimization (RTM)—using the Elba-93 sea trial dataset. Error metrics and processing complexities are systematically evaluated across stationary and maneuvering target scenarios. Rigorous non-parametric statistical tests reveal distinct operational boundaries: under stationary conditions dominated by systemic environmental mismatch, energy-based processors guarantee reliable baseline stability. Conversely, under snapshot-deficient dynamic conditions tracking a receding target, standard high-resolution subspace methods become highly vulnerable to trajectory jumps. In such highly dynamic scenarios, adaptive energy-based processors (specifically MVDR) exhibit the most stable tracking continuity and lowest numerical peak errors. Simultaneously, the operational adaptability of subspace methods is improved via covariance matrix reconstruction (CMR). Specifically, the RCM technique effectively decouples unstructured sensor noise, mitigating maximum trajectory deviations and providing a balanced trade-off between computational efficiency and robustness. Statistical evaluations confirm the fundamental performance boundaries in static environments, while highlighting sample-size limitations in highly dynamic scenarios, thereby establishing a realistic, evidence-based benchmark for marine engineering applications.

Matched-field processing (MFP) is a conventional method for underwater acoustic target localization but is often sensitive to systematic environmental mismatch. This paper presents a transformer-based deep learning framework for joint two-dimensional (2D) range and depth localization of a fixed underwater sound source. Using the Elba-93 sea trial dataset, the model is trained on synthetic acoustic data generated by the KRAKEN propagation model and evaluated on experimental data from a 48-element vertical line array. The proposed model is compared against MFP baselines for Bartlett, minimum variance distortionless response, robust sub-array minimum variance distortionless response (SA-MVDR), and neural network baselines for one-dimensional convolutional neural network (1D-CNN) and multi-layer perceptron (MLP). Results demonstrate that the transformer achieved a 2D root mean square error of 17.59 m, which reflects the repeatability and residual bias at this fixed, on-grid location under one specific simulation-to-experiment mismatch condition. Under the same fixed-position test condition, the transformer produced lower repeated-estimate errors than the 1D-CNN and MLP baselines. Furthermore, spatial sparsity analysis reveals that the model maintains consistent accuracy using input from a single hydrophone during the testing phase. Independent evaluations indicate that the transformer encodes full-array spatial priors into its neural weights during synthetic pre-training, enabling the transfer of spatial array gain to support single sensor deployments and offering a feasible pathway for simplified underwater surveillance systems.