A lightweight hybrid attention network with multi-scale feature integration for intelligent recognition of underwater acoustic targets Authors: Nahid-Al Mahmud, Tao Zhang, Yasir Iqbal, Farhana Bari Sumona, Anjum Iqbal, Yanzhang Geng, Wajid Khan, Basem Abu Zneid, Ievgen Zaitsev Published: 2026 ABSTRACT AbstractUnderwater Acoustic Target Recognition (UATR) plays a significant role in intelligent sonar and underwater environmental monitoring systems, but successful recognition in underwater environments of complex scenarios is a significant challenge because of the effects of ambient noise, signal attenuation, and dynamic propagation. The recent developments in deep learning have enhanced automatic generation of features in acoustic signal classification, and the vast majority of existing models represent a trade-off between recognition accuracy and computational efficiency. In this regard we introduce Lightweight Hybrid Attention Network with Multi Scale Feature Integration (DCAT) a new deep learning system that combines depthwise separable convolutions to extract local features efficiently in parallel in conjunction with transformer-based global temporal dependency modeling attention modules. The key innovation of DCAT is its adaptive fusion mechanism, which conditionally combines contextual information of the two transformer branches with different receptive fields with varying scales with the purpose of allowing the model to ideally capture small-scale local features and long-range acoustic structures. The framework employs robust preprocessing and feature engineering, including Zero Crossing Rate (ZCR), Root Mean Square Energy (RMSE), Mel-Frequency Cepstral Coefficients (MFCC), and Chroma features, computed from 22.05 kHz sampled audio, combined with data augmentation techniques—pitch shifting, time stretching, and Gaussian noise addition—to enhance generalization to real-world acoustic variability. Evaluated on two benchmark datasets, DeepShip and ShipsEar, DCAT achieves superior classification accuracies of 98.84% and 99.16%, respectively, while maintaining extremely low computational complexity of only 0.52 million parameters and 6.1 million FLOPs, supporting real-time inference with latency below 0.7 ms per sample. Comparative studies show that DCAT is more accurate and more efficient compared to state-of-the-art networks, including Transformer, ResNet1D, and AResNet, which substantiate its ability to trade-off discriminative power with resource economy. The suggested model sets a new standard of performance efficiency of underwater acoustic target recognition, which offers a promising framework to the next-generation autonomous sonar, underwater surveillance, and marine ecological monitoring systems.Keywords: Adaptive attention mechanism, Deep learning, Feature fusion, Multi-scale modeling, Sonar signal processing, Transformer network, Underwater acoustic recognition, Underwater target classificationSubject terms: Engineering, Mathematics and computing, Ocean sciences Introduction The recognition of underwater targets is a critical component of acoustic signal processing, and it offers the fundamental aspect of the intelligent sonar systems development. Although this topic has been the focus of research in underwater acoustic over the decades, it has continued to be significantly challenging due to the inherent complexities. Acoustic waves are chosen to carry out underwater detection and identification provisions because they propagate long distance in water1. Proper Underwater Target Recognition (UATR) is essential toward the situation awareness and the use of countermeasures in the sea. Therefore, the capacity to identify acoustic objectives (UATR) is basic to sonar activity, allowing possible target resolution underwater. The importance of the method is the broad application, including the monitoring ship-generated noise and assessing ecological impacts. It is also contributes significantly to the development of ship decommissioning strategies2. Human exploitation of marine life is at an unsustainable rate, and marine ecosystem are degrading faster than human life. According to scientific studies, marine ecosystem diversity and organism population have been significantly reduced by commercial shipping noise pollution3. The traditional UATR systems still face significant limitations. These issues involve dynamic target signatures, ambient-noise corruption, false returns from non-target objects. Another complication is the fact that the signals become attenuated due to the propagation effects4–8. This current advance of AI technology and wider access to publicly accessible underwater sound databases have allowed researchers to successfully apply deep learning models. These Machine learning models establish direct relationships between inputs and outputs through autonomous learning of complex nonlinear relationships. This functionality offers novel solutions to the chronic UATR issues9,10. In contemporary applications, scholars usually convert crude acoustic data into graphical time-frequency images. This is then followed by using CNN architectures to directly obtain classification features of these graphical inputs. STFT creates basic linear time-frequency plots, known as LOFAR spectrograms, which are fundamental in underwater acoustic processing. On the other hand, the Mel-scale transforms were used to introduce nonlinear warping to better characterize the energy distribution properties of the signal11–13. Residual and densely connected networks (ResNet, DenseNet) are better in the UATR application, when using Mel frequency spectrograms. This is interesting even when using classic architectures such as VGG19 do well in typical computer vision problems14,15. The Hongs system integrates a feature multi-scale feature maps into a personalized ResNet 18 framework. It also incorporates optimization of embedding layer and center loss. In this method, a greater underwater acoustic target detection capability is attained16. The time-domain feature extraction method was used in the research group headed by Meng. This approach was combined with SVM pattern recognition of multi class identification of underwater ship noises17. Scholars led by Shen developed an expert CNN framework. This design incorporates psychoacoustic processing units to enhance ship identification performance18. Lius group came up with an innovative model that combines volumetric time frequency representations with deep convolutional-recurrent learning. It was meant to be used in classifying sound in marine vessels with the help of the Ships Ear data2. The research team by Hong came up with a modified ResNet structure that has advanced feature extraction. This model was used to classify at 94.3% in their UATR experiment. This has been achieved based on the Ships Ear underwater acoustic dataset16. The solution of Luos RBM was experimented in UATR with Ships Ear dataset. It had a significant rate of accuracy of 93.17. Its underwater acoustic recognition tasks are highly effective as illustrated in this performance19. A new CNN-ELM framework was suggested by Hus research team. The model achieved 93.04% accuracy in the identification of actual civil ship noise when used on actual civil ship noise recordings. This shows that it is very effective when applied with real-life data20. The transfer learning model proposed by Ke et al. integrates RSSD for high-resonance feature extraction with a CRNNs architecture. This approach was aimed at processing three dimensional features of ship noise. It is an effective method when dealing with big data. It however performs poorly with small data. This weakness is mainly due to its rigorous pretraining procedure2,21. Deep learning has catalyzed progress in various fields, including foundational applications like computer vision22–25, detection of object26–28, language processing naturally29 and Speech enhancement30. As a result, researchers have developed diverse deep learning process to classify underwater targets19,31–34. The ability of deep neural networks to represent features in a very powerful way contributes to this development. Such abilities have made DNNs the new center of interest in UATR studies. Originally, UATR was performed through human identification of acoustic signals. Normally, the signals would be listened to the analysts as raw recordings or preprocessed audio files in order to categorize targets. This method, however, has a number of apparent limitations and drawbacks. The duration of the manual analysis and effort increases with the increased amount of data. Therefore, it becomes hard to attain real time recognition. More importantly, underwater acoustic signals have a large frequency range with a significant part of it being outside of the human auditory range. The long-term consequences of experiencing these abnormal acoustic signals may be destructive to human health. Various spectral analysis techniques have been applied to identify the target, such as power spectrum examination, LOFAR and DEMON. Utilizing both the frequency and time-frequency domains, these techniques extract the discriminative techniques of acoustic signals in underwater35,36. The process of GFCC and MFCC are still popular in underwater acoustic signal analysis because they can characterize the spectral in an effective way37. Scientists have come up with hybrid systems that combine machine learning and signal processing methodology. These systems convert the acoustic frequency characteristics into predictive inputs of automated recognition38. The powerful learning capacity of deep neural networks enables the autonomous extraction hierarchical features from signals. Consequently, recent approaches have successfully use direct raw signal processing for underwater target recognition39. In conventional deep learning (e.g., image recognition) applications, raw pixel data is fed directly into neural networks. However, the extended durations and high sampling rates of hydroacoustic signals require pre-processing such as frame segmentation. The direct input of raw data risks overwhelming the model while introducing noise and interference40. Hu architected a neural network to process underwater acoustic signals directly. His design features depthwise separable convolution and time-discrete convolution components specifically tailored for UATR tasks41. Li introduced ASTEM DCNN, a timbre-sensitive deep convolution neural network. It performs line spectrum extraction and frequency component fusion to achieve more interpretable feature representation42. Song introduced a hybrid deep learning approach for UATR. It blends a 1D CNN for extracting spatial features with an LSTM network for modeling time based sequence43. The method of Yangs involves feature extraction of the signal of ship noise and formation of weight learning neurons. It manages to project important underwater target parameters such as the size of the vessel and its state of operation44. As a core component of acoustic signal processing, underwater target recognition enables the innovation of intelligent underwater systems. Marine acoustics Underwater acoustics has long been a focus of marine research. Nevertheless, it is complex enough to be a significant technical challenge. Transformer architectures are included in recent research on UATR in order to capture long-range dependencies in hydroacoustic signals. This method has been more successful in capturing global spectral temporal contexts as compared to the traditional CNNs45. Nonetheless, conventional transformers require computational requirements that are prohibitive, because of quadratic self-attention complexity. Such a costly price does not allow real-time processing of large acoustic sequences46,47. Depthwise separable convolutions provide a resolution by dividing up the operation. This minimizes parameters and enhances computational efficiency that is not compromised on performance48,49. In this paper, it is suggested that we can use Depthwise Separable Convolutional Adaptive Transformer (DCAT). It is a novel architecture that is used in the Underwater Acoustic Target Recognition (UATR). This architecture integrates effective depthwise separable convolutions with local and global characteristics of adaptive transformers attention mechanisms of sequential acoustic data. The depthwise separable convolutions involve the use of depthwise separable convolutions that have a considerable reduction of the parameter but which preserve the high performance of extraction of features. The architecture uses parallel transformer branches, each having varying attention head configuration. The design allows the multi scale temporal pattern learning to be particularly effective in balancing the local processing of convolutions with a global context modeling of transformers. Besides, adaptive fusion module smartly integrates the results of both branches of the transformer, adjusting their weight based on context demands. The adaptive fusion strategy optimizes classification accuracy through dynamic classification of most discriminative features. The model uses extracted features including Root Mean Square Energy (RMSE), Zero Crossing Rate (ZCR), and Mel-frequency Cepstral Coefficients (MFCC) for classification. The rest of the paper is structured as follows: Sect. 2 outlines the technique of the proposed deep learning approach, Sect. 3 explains the model architecture and training process, Sect. 4 summarized the experimental setup, Sect. 5 interprets the analysis and findings of the experiment, and the study was concluded in Sect. 6. Methodology Data preprocessing The initial stage of the proposed UATR method involves the processing of raw underwater acoustic signals, represented as , where each ​ is an audio sample. These records are first sampled at 32 kHz (DeepShip dataset) and 52.7 kHz ( ShipsEar dataset ). But in the name of uniformity and compatibility between datasets, we resampled all audio samples to an identical frequency of 22.05 kHz, a common step in minimizing computational costs without loss of enough temporal resolution to perform recognition tasks. The audio samples after resampling are split into fixed length frames where each frame is 5 s long. An audio sample of less than this fixed duration is zero-padded; an audio sample of more than this fixed length is truncated. Each segment is made to have a fixed length of time steps, ensuring uniformity across the dataset: In this case, is the constant duration of time intervals, which is the product of the sampling rate SR with the constant duration , i.e., in which = 22,050 samples per second (22.05 kHz), and seconds = 5 s. The in operation indicates the operation of zero expanding or truncating the sound signal to the desired length . Data augmentation After preprocessing, data augmentation is applied to artificially increase the training dataset that improves the generalization of models ability as illustrated in Fig. 1. The data that has been artificially varied in training facilitates more robust feature learning that leads to enhanced performance in actual underwater acoustic variation. The augmentation operation produces two new versions of every initial sample that has been modified, and the overall volume of training data is tripled. All augmentation methods were applied to the training set only and all random processes were controlled with a random seed of 42 in order to reproduce them. The augmentation methods used in this method are as follows: Overview of the proposed DCAT model methodology for Underwater Acoustic Target Recognition (UTAR), including data preprocessing, feature extraction, data splitting, and the flow of data through the model. Pitch shifting The parameter p scales the frequency spectrum causing a controlled change in the pitch of the acoustic sample. The pitch shifting equation is as follows: In this case t is the time variable and is used to manipulate the extent of pitch change. The parameter is limited strictly to a small range (± 2 semitones) to give a realistic simulation of the Doppler shifts without introducing artificial classes. Consequently, the model can generalize to diverse real-world acoustic patterns while maintaining the discriminative features of the target class. Time stretching The method does not modify the pitch but modifies the length of the signal and replicates changes in speed underwater. The time-stretching operation can be represented mathematically as: Time domains of a signal is denoted as and transformation as is used to manipulate both the duration. The scaling factor is maintained within a range of 0.8 to 1.2 because it is necessary to maintain physical plausibility of variations but retain class identity. This increases precision in target recognition when moving, and this mimics the acoustic signatures of the targets under varying velocity conditions. Through this, the model becomes more resistant to speed related changes in acoustic signature of the targets without the need to form new classes. Noise addition To strongly model realistic ambient ocean noise we combine the signals with colored noise with spectral profile that decreases with frequency, which is a close approximation of other spectral profile models such as pink noise ( noise). This method works better in physical terms than white Gaussian noise in underwater conditions where the lower frequencies dominate as it is observed in oceanic noise sources, which include marine life, engine hums, and ocean currents. The signal with noise is mathematically represented as: Where, represents a colored noise signal with a spectral characteristic (with for pink noise), and α is the amplitude scaling coefficient, calibrated to real signal-to-noise ratios (SNRs) in the range of 10–20 dB. By commanding the model to learn on acoustically realistic, perturbed environments that are close to realistic underwater environments, this noise augmentation technique increases the robustness of the model. Augmentation strategies are applied in the training phase to develop a richer dataset that enhances its realism and quality. The larger and more extensive training data set exposes the model to more acoustic variations and this improves and enhances the ability to generalize and perform effectively in real-world UATR scenarios. Feature extraction Once the data has been augmented, the acoustic signals of the underwater environment are extracted to feature the major aspects of those signals. These characteristics are Zero Crossing Rate (ZCR), MFCC, Root Mean Square Energy (RMSE), and Chroma is displayed in Fig. 2 that represent the spectral and temporal characteristics of the raw waveform signals which is critical towards the detection of acoustic targets. Time domain characteristics are given a high priority because they are more resistant to ambient noise and computationally efficient (ZCR and RMSE). The characteristics are specified in subsequent subsections. Feature extraction pipeline for proposed DCAT model for UTAR showing transformation of (a) raw audio from DeepShip dataset of 5 s passenger signal sample into: (b) MFCC coefficients, (b) Chroma features, (c) zero crossing rate, and (d) RMS energy. These features were computed using 2048-sample frames with 512-sample hop length, with statistics (mean and standard deviation) calculated across frames, resulting in a 55-dimensional feature vector per audio sample. Zero crossing rate (ZCR) Zero Crossing Rate (ZCR) is used to identify the times number of signal passes through the zero point per frame, indicating the Sign change rate. This aspect assists in describing the content of the frequency signals. For a signal x, the ZCR is computed as: Where is the indicator variable, which is 1 when and ​ is negative, and 0 otherwise. Although time domain features, including ZCR, have proven to be resistant to ambient noise and easy to compute, the performance of ZCRs is compromised with ship noise, which possesses a combination of broadband and tonal coherence. Under such circumstances ZCR might fail to fully record the variations of broadband noise. Nonetheless it is useful in the identification of transients and sudden frequency variations in the signal. ZCR, when added to other characteristics, such as the RMSE, MFCC, and Chroma, helps to make the underwater acoustic environment more comprehensive and effective. Processing ZCR two features mean and standard deviation . Root mean square energy (RMSE) RMSE is used to define the effective magnitude of a signal through calculating the energy-equivalent constant amplitude of the signal. RMSE is mathematically defined as the square root of the arithmetic mean of squared instantaneous amplitudes of signal: RMSE is a time-dependent feature which is less affected by ambient noise and signal distortions than frequency domain features, provides a stronger description of the magnitude of the signal and guarantees reliable target detection in noisy underwater conditions. In this respect, time-domain features are more desirable than frequency or time-frequency representations as they are simpler and more resilient to noise. Two features mean and standard deviation were obtained. Mel-frequency cepstral coefficients (MFCC) MFCCs are commonly used as normal form of feature representation of spectral properties of audio signals in speech and sound processing systems. The MFCCs are computed to obtain spectral characteristics of the signal, and they resemble the way the human auditory system perceives frequencies. The MFCCs are computed as: The Discrete Cosine Transform (DCT) works on the log-transformed results of the Mel-scaled filter bank, which results in the final cepstral coefficients. Even though MFCCs are commonly applied in speech processing, they also be useful in UATR, such as ship noise, since they extract salient spectral characteristics that are used to differentiate between acoustic sources. Although there are various disparities between human hearing and ship sound, MFCCs have the ability to reflect the underlying spectral character of the signal especially in tonal constituents and periodicity in ship engine sound. MFCCs are also capable of capturing the appropriate features of discriminating between ship noise and other sounds in the environment by concentrating on the Mel scale, which focuses on the lower frequencies. From MFCC, we compute 26 features including 13 MFCC coefficients, each with mean and standard deviation, and total of 26 features. Chroma By mapping spectral energy to musical pitch classes, Chroma features extract perceptually meaningful harmonic information. It is defined as: Chroma features are also useful in ship noise analysis and are typically used in music and audio processing. Although originally used to represent musical pitch classes, Chroma is useful in representing the harmonic content and periodic variations in mechanical sounds, e.g. engine hums and propeller noises. It has the assistance of distinguishing tonal contents of ship noise and assists in differentiating the types of ships or their state of operation. From Chroma, we extract 24 features, 12 Chroma bins (C, C#, D, D#, E, F, F#, G, G#, A, A#, B ), each with mean and standard deviation . Features vector and computation The above feature extractions were computed using frame-based analysis with 2048-sample frames and 512-sample hop length. For each 5-second audio segment (110,250 samples at 22,050 Hz), this results in approximately 215 frames. Statistical moments (mean and standard deviation) were calculated across all frames for each feature, yielding a fixed-dimensional feature vector per audio sample. A 55-D feature vector is constructed for each audio clip by computing the standard deviation (σ) and temporal mean (µ) of all features. Specifically, for a sample ​, the feature vector is given by: Here, µ denotes the arithmetic average while σ measures the dispersion of feature values. Using Standard Scaler, features are centered (zero mean) and scaled (unit variance) to create a normalized representation: X denotes the feature matrix, while m encodes modality information (e.g., specifying the data source as ‘ShipAudio’). Model architecture: depthwise convolutional adaptive transformer (DCAT) Depthwise Convolution Adaptive Transformer (DCAT) is a novel combination of depthwise separable convolution operations and transformer-like layers of attentions, the structure of which is an effective architecture in processing sequential data as shown in Fig. 3. This architecture is able to use local feature extraction and global context modeling to perform very well in time sensitive tasks such as sequence classification and forecasting. Architecture of Depthwise Separable Convolutional Adaptive Transformer (DCAT) for Underwater Acoustic Target Recognition. Input layer The model is initialized with an input sequence, that indicates one-dimensional sequential data with 55 features. Being the first layer to process it, the layer encodes each point of the input sequence with each feature. The model architecture is highly optimized to work with sequential/ temporal data, which is why it can be applicable in analyzing sensor data and signal processing. Depthwise separable convolution stack The model uses sequences of depthwise separable convolution layers with the aim of extracting significant features in the raw input data. A depthwise separable version of convolutions is a cost-efficient variant of the convolutions that can replace the regular convolutions reducing parameters and computational load by orders of magnitude, which is particularly beneficial for models with many layers. This efficiency has been recently demonstrated in audio processing tasks, such as speech enhancement50 and audio super-resolution51, where depthwise separable convolutions enabled compact, high-performing models suitable for real-time execution. The operation initiates with a depthwise convolution stage, where independent filters process each input channel separately. At layer iii, the depthwise convolution yields the following output: Here, ​ represents the output from the previous layer, and refers to the weights of the depthwise convolution for the block. The kernel size is set to 3, and the strides for each block are defined as (1,2,2,2), reducing the spatial dimension of the feature map progressively. Next, a pointwise convolution is applied: Where, is the output from the depthwise convolution, and represents the pointwise convolution weights. The pointwise convolution allows the network to learn cross-channel dependencies. After the convolutional layers, batch normalization and PReLU activation are applied to the outputs, ensuring stable training and introducing non-linearity: After several convolutional blocks, the final output after the last convolutional block is, where the spatial dimension has been reduced, and the depth of the feature map has increased. Parallel transformer branches The next stage involves processing the extracted features through two parallel transformer branches. Both branches take as their input and to model long-range dependencies in the sequence, multi-head attention mechanisms are applied. In Branch 1, the model applies a multi-head attention (MHA) mechanism with 4 attention heads: Where represents the multi-head attention with 4 heads. This attention layer computes the pairwise relationships between the features, allowing the model to focus on various parts of sequence. The output from the attention mechanism is first enhanced with a residual connection and then normalized by a Layer Normalization step: The normalized output is passed through a feed-forward network (FFN): Where , and are the weights and biases of the FFN. A second residual connection is added, followed by Layer Normalization: In Branch 2, the operations are identical but with 8 attention heads: The output of Branch 2 undergoes similar processing, this branch follows the same processing as Branch 1 and outputs the final sequence representation is . After the attention mechanisms, both branches undergo pooling. Branch 1 uses Global Average Pooling (GAP) to produce a fixed-size vector , while Branch 2 uses Global Max Pooling (GMP) to produce : The Global Average Pooling (GAP) is applied and its operation is expressed as: The Global Max Pooling (GMP) is applied to branch2 and it is computed as: This results in two feature vectors, ​ and ​​, both of size. Adaptive fusion block In the adaptive fusion block, the outputs from the two transformer branches are combined using attention mechanisms to adaptively weigh the contributions of each branch. The pooled feature vectors ​ and ​ are reshaped into sequences with a single time step: These reshaped vectors are then concatenated along the sequence dimension: The concatenated features are processed by another multi-head attention layer: Thus, dependencies between the branches are captured by the model. The outputs attention mechanism are then concatenated with the original features: After concatenation, Layer Normalization is applied: The attention scores are computed using a dense layer with sigmoid activation: Where denotes the sigmoid activation function. The Softmax operation is applied to the attention scores: These attention weights are then used to weight the features: Where represents multiplication applied element by element. The final fused features are obtained by summing the weighted features: Classifier The fused feature vector is passed through two fully connected layers for classification. The first dense layer uses a PReLU activation: After this, dropout is applied to prevent overfitting, with a rate of 0.5. The second dense layer also uses ReLU activation: Dropout with a rate of 0.3 is applied after the second layer. The final output is obtained through the output layer: The Softmax activation offers a probability distribution among the four classes, which allows the model to categorize more than four classes. Model training Depthwise Convolutional Adaptive Transformer (DCAT) is a model in which Categorical Cross-Entropy Loss is used to reduce disparity between output of the predicted value ​ and actual labels ​. The loss can be computed as follows: Where is the number of classes (4), ​ is the true label of (one-hot encoded), and is the predicted probability of the class. The predicted probabilities ​, are subjected to the logarithm and not the actual labels ​. The training process makes use of the Adam optimizer which can tune the learning rate to implement a high-performance training whereby the DCAT model will learn to determine the probabilities of the classes of the output of each input sequence. This procedure enhances the recognition capabilities of the models of the underwater acoustic targets in different conditions. Training setups Data collection Two well-known and publicly available datasets of underwater acoustic targets (DeepShip10 and ShipsEar9 (both targeting ship radiated noise) were measured to conduct the experiment. In this paper, we shall discuss the method of this dataset in UATR. The description for both data sets follows. DeepShip was published by the Canadian Ocean Observatory Network and contains ship noise measurements collected around the Delta Node in Georgia on 2 May 2016 through to 4 October 2018. The information is useful as it represents an actual image of the state of the marine environment at various sea conditions and seasons. The logged information does not only contain ship signals but also natural background noise, marine mammal sounds as well as other noises of human activity. The data set includes four varieties of vessels namely tugboats, oil tankers, cargo ships and passenger ships, which contain 265 dissimilar underwater recordings of unquestioned durations that range 47 h and 4 min. In this study, the data is analyzed in the form of 5 s segments. This was achieved without any repetitions in order to avoid overlapping and a total of 10,000 audio clips were obtained. ShipsEar dataset is an open access underwater ship noise dataset recorded in a total of approximately 3 h. These were recorded in the natural marine ecosystem in the Atlantic coast of northwestern Spain, that is, in the Vigo Valley and near the port of Vigo. The data was collected in the fall 2012 and the summer of 2013. The materials are 90 WAV files, where the length of the recording is between 15 s and 10 min. These tapes contain sounds of 11 classes of ships, which are categorized in four classes of ship radiated noise in terms of the size and nature of the ship. In order to make analysis easier, the recordings were separated into 5-second sound files having no overlap between two successive files. This gave a total of 1,440 clips. To analyze them, the recordings were re-arranged into five groups depending on the sizes of the vessels, which were Class A, B, C, D, and E. In order to create the datasets, the recording-wise division was used, so that each recording of each type (tugboats, oil tankers, etc.) was included in only one set (i.e., training, validation, or testing sets) and did not overlap with any other set. For the ShipsEar dataset, an 8:1:1 split was used, allocating 80% of the data for training and 10% each for validation and testing. In the case of DeepShip, the data was divided into the following proportions, 7:1.5:1.5 whereby 70% consisted of the training data, 15% was the validation data, and the remaining 15% consisted of test data, as in Table 1. This division, as represented in recording is such that no frames of the same recording are represented in the training and the testing sets, which is a practical way of avoiding data leakage. Table 1Dataset distribution across training, validation and testing set (with fixed frames of 5 s).DatasetsOriginal recordingsTotal clipsTraining setValidation set (30%)Testing setDurationShipsEar901,4401,152 (80% of total)144 (10% of total)144 (10% of total)5sDeepShip26510,0007,000 (70%of total)1500 (15% of total)1500 (15% of total)5s Dataset distribution across training, validation and testing set (with fixed frames of 5 s). Training parameters The model development (TensorFlow) is based on the deep learning framework in this research. The model applied is Depthwise Separable Convolutional Adaptable Transformer of Underwater Acoustic Target Recognition (DCAT). During training, the model parameters will be optimized via the Adam optimizer, set to a learning rate of 0.001. Optimization with 64-sample mini batches and categorical cross-entropy loss is minimized. Training is done with Adam and moment decay rates of 0.9 (first-order) and 0.999 (second-order) to obtain better generalization and optimization stability. Architecture is an ensemble of depthwise separable convolutional layers with transformer blocks to allow efficient feature extraction and representation learning. The systematic frame-based analysis method with hard-parameter settings was applied to extract the features in order to make the results consistent and reproducible across all experiments. The audio samples were cut at a constant 5.0 s and sampled at 22,050 Hz. Frame-level analysis was done with 2048-sample frames (~ 93 ms) and a 512-sample hop length (~ 23 ms), a Hann window function and a 2048 sample FFT size. To extract spectral features, 128 Mel filter banks were involved, and 13 MFCC coefficients were calculated to indicate the spectral envelope. Chroma features were extracted using 12 pitch class bins (C, C#, D, D#, E, F, F#, G, G#, A, A#, B) to capture harmonic content. To prevent data leakage, the datasets were split at the recording level. Using a fixed random seed of 42 for reproducibility, DeepShip was divided into 70% training, 15% validation, and 15% testing, while ShipsEar was split into 80% training, 10% validation, and 10% testing. The application was based on librosa 0.10.0 that processed audio, scikit-learn 1.3.0 that normalized features with StandardScaler (zero mean, unit variance), and TensorFlow 2.12.0 that developed the model. All these variables were maintained at a constant during the experiment in order to compare directly and reproduce the results of the experiment. The training program is based on 30 epochs including early stopping and dynamic learning rate decay to improve generalization and model accuracy. Training was done in a dual-GPU architecture (NVIDIA RTX 2080 Ti) that had CUDA 12.2 (Driver: 535.171.04). The computing resources allow optimized throughput in batch processing and calculation of gradient during training. Evaluation metrics Model evaluation is based on standard classification metrics: Accuracy, Precision, Recall, and the F1-score. These metrics are necessary in high stakes application in which false positives or false negatives cannot be afforded, and all measures must be of excellence. The report presents the precision, recall, F1-score, and the number of test instances (support) of each target class. The evaluation employs confusion matrices, ROC curves, and Precision-Recall curves, providing a comprehensive visual analysis for deeper performance insight. Mathematical Expressions for these metrics are: Results and discussion Evaluation on classification metrics using the DeepShip dataset Training and validation trends Figure 4 illustrates the training and validation loss/accuracy of the DCAT model over 30 epochs on the DeepShip dataset, demonstrating the model’s capability to learn from underwater acoustic targets. Figure 4a shows that the training loss begins with about 0.5 and it is gradually declining to a minimum point at the fifth epoch. The test loss, which begins at a rate of about 5.0, rises and then falls and levels of around 0.1 by the 16th epoch. This variation of the test loss means that there is some initial challenge to the model but it levels as the training level advances. In Fig. 4b, the training accuracy starts with 0.8 but quickly builds to close to 1.0 at the end of the few epochs. The accuracy of the test follows a comparable trend where the accuracy at the beginning of the test is 0.1 and it increases to a range of about 98.84% at epoch 25 and it levels off. The slight difference between the training and test accuracy curve indicates that the generalization and convergence are good and the final test accuracy is outstanding at 98.84%. These results demonstrate the effectiveness of the proposed DCAT model in achieving accurate underwater acoustic target recognition. Representation of the model (a) Training and Test Loss (b) Accuracy for DCAT Model on the DeepShip Dataset. Confusion matrix The confusion matrix as shown in Fig. 5 depicts the performance of the model used on the DeepShip dataset, which is Depthwise Separable Convolutional Adaptive Transformer (DCAT) and that the model achieved a high performance in the classification of underwater targets, namely: Cargo, Passengership, Tanker and Tug. The confusion matrix displays class-wise accuracy percentages for clearer interpretation. The model achieves 96.2% accuracy for the Cargo class, with only 3.8% misclassified as Passengership. The model achieves 97.9% accuracy for Passengership detection, with only 2.1% misclassified as Cargo. The model has a perfect discrimination power of Tanker and Tug vessels indicating 100% classification with no errors. The confusion matrix confirms a highly accurate classification ability of the DCAT model with cargo and PassengerShip being the only two categories which are a little confused. This brings a fore to the classification strength of the system in object detection of different underwater acoustic targets. Confusion matrix for the DCAT model on the DeepShip dataset classification. Receiver operating characteristic (ROC) curve Figure 6 illustrates the ROC curves of the classification of the DCAT model with all classes and the Deepship dataset. The ROC plot illustrates individual ROC curves (one per class Cargo, Passengership, Tanker, and Tug) in addition, micro-average and macro-average performance curves. All the ROC curves achieve an optimal performance Area under the Curve (AUC) of 1.00, which indicates the flawlessness of discriminating power of the model on all vessel types. Micro-average ROC (dotted) and macro-average (dashed) show perfect average under AUC (1.00) confirming no true-positive/false-positive trade off and the model is optimally performing. The DCAT model proves to be significantly effective in classification of all DeepShip classes with exceptional accuracy in the placement of underwater targets recognition. ROC curve for proposed DCAT model for UTAR classification on DeepShip dataset. t-SNE plots The t-SNE52 is used to evaluate the proposed model which transfer high dimensional data into lower dimensional space to visualized in clusters. Figure 7 shows the results of the t-SEN plot of the feature embeddings learned by the DCAT model of the DeepShip dataset, exposing the model behavior in clustering. The 2D t-SNE projection shows how the point that represent the data of four DeepShip classes (Cargo, Passengership, Tanker, Tug) are clustered. The t-SNE plot shows a high range of separation between classes, and the Tanker and the Tug classes that can be described as being compact and isolated. The visualization shows partial overlap between Cargo and Passengership clusters, with some Cargo instances encroaching on PassengerShip’s feature space. The limited Cargo/Passengership overlap has negligible impact, as evidenced by the model’s near-perfect 98.84% accuracy on test data. Most of the classes are separated on the basis of features very effectively with a small amount of Cargo/Passengership overlap as shown by the DCAT model, which is effective in underwater acoustic classification. t-SNE visualization for the DCAT model on the DeepShip dataset for UTAR. Precision-recall curve Figure 8 constitutes a Precision-Recall (PR) curve that is measured the performance of DCAT model on DeepShip dataset in classification. The plot compares precision-recall tradeoffs on an individual class (drawn in different colors) to the results of the micro-average. The Precision-Recall curve shows the predictive performance of the model at various levels of decision threshold. The Cargo, Tanker, Tug classes achieve an ideal Average Precision (AP = 1.00) which indicates that they have no accuracy-recall trade off in classifying the data. The Passengership class presents almost perfect average precision (AP = 0.99) and good classification performance even in case deviations occur. Combining all of the categories, the model yields perfect micro-average precision (AP = 1.00), which is a strong indicating of uniform high performance. The best F1 score (0.98) happens when precision (P) and recall (R) have equal values of 0.98, which can be considered close to the perfect balance of accuracy and coverage. The outcomes prove a rather high precision and recall rates on all of the classifications, which proves the model can be applicable to underwater target classification in the DeepShip dataset. The DCAT model proves to have excellent classification with its reported precision and recall being very high in all the target classes. Which is prove that it is effective in underwater acoustic recognition tasks. Precision-recall curve for DCAT model on the DeepShip dataset. Performance evaluation on objective metrics using DeepShip dataset Table 2 illustrates the systematic performance of DCAT, including class-wise precision, recall, and F1-scores as well as the overall accuracy on DeepShip benchmark. The model has a Cargo classification precision of 98.26%, which implies that almost all of its positive predictions in this category are correct. The 0.9725 recall score demonstrates the model’s high sensitivity, correctly identifying nearly all genuine Cargo cases. The harmonic mean (F1 = 0.9775) confirms consistently high performance for Cargo, with both precision and recall contributing equally. In Passengership identification, the model records 97.29% precision, meaning that it has strong positive predictions in this category. A recall value of 0.9829 shows how the model has found 98.29 of total true Passengership cases in the data. The F1-score of 0.9779 shows a close-to-perfect balance between the recall and precision, where there is slight skewing towards the sensitivity of detection. There is also good accuracy in the model with regards to the classification of Tankers (0.9959) with the correct prediction where the accuracy of prediction stands at 99.59%. The model also fits all Tanker cases and achieves an almost perfect F1-score (0.9979) and perfect recall (1.0000), which is an excellent overall performance in classification. With the model, in Tug classification, the model produces: (1) a perfect metric in precision (1.0000) and no false positives, (2) almost perfect metrics in recall (1) (0.9958), (3) colossal F1-score (0.9979), leading to yet another route to an excellent performance. The model was found to be 98.84, a fact that shows that the model had an excellent performance in the classification performance on the target categories. These findings affirm the excellent classification of the model in all the target groups in the dataset. The DCAT model demonstrates great classification accuracy with high precision, recall, and F1-score using all types of vessels and a high general accuracy rate of 98.84%. It is proving that it is proficiency as an underwater acoustics recognition scheme. Table 2DCAT model’s Performance evaluation on objective metrics.ClassesPrecisionRecallF1-ScoreCargo0.9826320.97250.977539PassengerShip0.9728870.9829170.977876Tanker0.99585110.997921Tug10.9958330.997912Accuracy0.9884130.9884130.988413 DCAT model’s Performance evaluation on objective metrics. Table 3 contrasts the performance of DCAT model with various feature extraction methods such as individual methods (ZCR, RMSE, MFCC, Chroma) and the combination of methods. With Zero Crossing Rate (ZCR) capabilities, the model achieves an 91.43% accuracy, 92.64% recall, 91.49 F1-score, and overall, 91.74 accuracy. The metrics indicate good Classification ability, where the recall (92.64) is slightly higher than the Precision (91.43), which means that the system has a slightly higher detection sensitivity than prediction certainty. The results obtain with the use of RMSE features are slightly better: P = 91.79, R = 91.84, F1 = 91.77, Accuracy = 91.82 - the precision recall equilibrium is better than with ZCR. With MFCC features, the model attains better results: Precision = 93.01, Recall = 92.51, F1-Score = 92.43 and Accuracy = 92.49. The findings substantiate the hypothesis that MFCC features are better than the ZCR and RMSE to improve the accuracy of classification in this underwater acoustic recognition task. Chroma features are also comparable in terms of performance to MFCCs with P = 93.47, R = 92.39, F1 = 92.34 and Accuracy = 92.46 indicating that they are highly suitable in underwater target identification. Precision, Recall, F1-Score, and Accuracy are all 98.84 representing a significant increase in classification capacity when ZCR, RMSE, MFCC, and Chroma features are combined. A combination of these complementary characteristics gives the most discriminative signal characterization, a feat that remains unmatched in classification accuracy. The results are conclusive that combining the features of ZCR, RMSE, MFCC, and Chroma boosts greatly the classification ability and the DCAT model can achieve 98.84% with respect to all the key performance indicators. These findings validate the substantial performance of the models in the context of underwater acoustic classification by using complementary features representation. Table 3Performance comparison of the proposed DCAT model with different feature extractors.ModelFeaturesPrecisionRecallF1-ScoreAccuracyProposed DCATZCR91.4392.6491.4991.74RMSE91.7991.8491.7791.82MFCC93.0192.5192.4392.49Chroma93.4792.3992.3492.46ZCR, RMSE, MFCC, Chroma98.8498.8498.8498.84 Performance comparison of the proposed DCAT model with different feature extractors. Table 4 assesses performance of the model and presents the proposed DCAT model by data augmentation using the DeepShip dataset. The model is significantly enhanced in classification performance with augmentation methods like pitch shifting, time stretching, and noise addition. The model using no augmentation attains a precision of 91.65, a recall of 91.64, F1-score of 91.71, and an accuracy of 91.65. Nevertheless, the introduction of augmentation results in the increment of these metrics to 98.84, which means that there is a significant increase in the generalization capacity of model and its performance in realistic underwater acoustic environments. The higher-level of training data represented provided by augmentation leads to a significantly better classification performance, which proves useful in the real world. Table 4Performance evaluation of the proposed DCAT model with and without augmentation using the DeepShip dataset.ModelAugmentationPrecisionRecallF1-ScoreAccuracyProposed DCMT✕91.6591.6491.7191.65✓98.8498.8498.8498.84 Performance evaluation of the proposed DCAT model with and without augmentation using the DeepShip dataset. Comparison against existing methods Table 5 gives the comparison of DCAT with state-of-the-art models on six metrics: accuracy, model size (Params), computational cost (FLOPs), and inference speed (RT), features used, and training time (h). DCAT has a highest accuracy of 98.84% in underwater acoustic target recognition, compared to all the benchmarked models. Using the lightweight DCATs architecture (0.52 M parameters, 6.1 M floating point operations) allows inference (0.7ms) with a moderate training time (47.2 h). The findings show the best balance between high accuracy and low computation overhead of DCATs. Also, DCAT shows good results when using alternative feature extraction methods, including FFT (98.25% accuracy) and MS_STFT+LE+FC (98.06% accuracy), which further illustrates its flexibility and effectiveness in the various configurations, and also out-performs baselines, including ResNet1D_S and AResnet. Transformer model achieves 95.3% accuracy but requires more resources—it uses 2.55 M parameters and 4.3 M FLOPs, and it takes 13.8 h of training. With the highest 99% accuracy, AResnet requires significantly more resources—it requires 9.47 M parameters and 1460 M FLOPS, and it takes 15.5 h to train. Although it is relatively simple 82.9% accurate, MSLEFC is computationally intensive (15.75 M parameters, 1556.9 M FLOPs) and inefficient in terms of feature representation and has long training durations. With 98.42% accuracy, ResNet1D_B is efficient deep learning with 2.1 M parameters (4× thinner than AResnet), 13.3 M FLOPs, sub-milliseconds latency (1.2ms), and conventional 47.2 h training. The lightweight ResNet1D_S model (95.72% accuracy, 0.17 M parameters and 6.8 M FLOPs) is a very resource-consuming model. Both versions of the ResNet are efficient but require a substantial amount of time for training. DCAT is more accurate, consumes fewer resources (parameters/FLOPs) and has a realistic training time (47.2 h) than current models. DCAT position as a computationally optimal underlying architecture in the classification of underwater acoustics. Table 5DCAT model performance comparison against existing methods using DeepShip dataset.ModelsAccuracyParamsFlops [M]RT (ms)FeaturesDuration (h)Transformer5395.32.554.3LogMel13.8AResnet54999.471460LM+MFCC+ CCTZ15.5MSLEFC5582.915.751556.9MS_STFT + LE+FC47.2End2end model5691.077.2None47.2Hybrid model5697.2720.6MFCC47.2ResNet1D_S5795.72/95.3 ± 0.280.176.80.8FFT47.2ResNet1D_B5798.42/98.1 ± 0.192.113.31.2FFT47.2Proposed DCAT98.25/98.17 ± 0.160.526.10.7FFT47.2Proposed DCAT98.06/97.98 ± 0.230.526.10.7MS_STFT + LE+FC47.2 Proposed DCAT 98.84/98.51 ± 0.13 0.52 6.1 0.7 ZCR, RMSE, MFCC, Chroma47.2Significant values are in bold. DCAT model performance comparison against existing methods using DeepShip dataset. Significant values are in bold. Additional experiments on ShipsEar dataset Table 6 compares DCAT with baseline models based on the ShipsEar dataset and evaluates the performance of both approaches in accuracy and model size (Params), computational cost (FLOPs), processing speed (RT), and feature composition. DCAT shows good performance of 99.16% accuracy with lean computational requirements of only 0.52 M parameters and 3.2 M FLOPs and rapid 0.6-minute time of execution. The combination of the ZCR, RMSE, MFCC, and Chroma features results in the optimal performance of the model without affecting computational efficiency. Also, DCAT demonstrates high performance with various standard sets of features: STFT (98.93/98.89 ± 0.16 accuracy), FFT (98.96/98.86 ± 0.13 accuracy), which exceeds the baselines. ResNet1D_S achieves 97.62% accuracy using only 0.17 M parameters and only 3.4 M FLOPs—lean computer needs—and has an effective runtime of only 0.7 min. ResNet1D_B model is highly practical (98.89% accuracy) with 2.1 million parameters, 5.0 million operations, and 66 s (1.1 m) runtime. Transformers (96.9%) and MR-CNN_A (98.87%) both have a competitive level of accuracy, but DCAT is more efficient - it provides high results with much fewer parameters and faster processing. Although AMNet-S has a marginally higher 99.4% accuracy than DCAT, it has a higher resource requirement and slower training. Overall, the Proposed DCAT model is an amazingly efficient and fastest performing solution of underwater acoustic target recognition that allows an amazing accuracy of 99.16% with minimal computational requirements and a processing time of only 0.6 min (RT1 m). The findings affirm that by assessing the proposed DCAT model using ZCR, RMSE, MFCC and Chroma features and by using standard STFT and FFT within the adaptive transformer framework, one obtains a well-tuned model that is optimized with ShipsEar dataset, and all baselines are outperformed. Table 6Performance comparison of the proposed DCAT model and existing methods on the ShipEar dataset.ModelsAccuracyParamsFlops [M]RT1 (m)FeaturesResNet2D181694.311.1918803.9LM+MFCC+CCTZSVM5894.490.1 20 3.4203.4DAWN + DRATransformer5396.92.55–4.3LogMelMR-CNN_A5998.874.35415.85–MFCCAMNet-N392.20.51140375STFTAMNet-T397.61.69170443STFTAMNet-S399.45.47370481STFTResNet1D_S5797.62/97.2 ± 0.440.173.40.7FFTResNet1D_B5798.89/98.5 ± 0.282.15.01.1FFTProposed DCAT98.93/98.89 ± 0.160.523.20.6STFTProposed DCAT98.96/98.86 ± 0.130.523.20.6FFTProposed DCAT99.16/99.09 ± 0.38 0.52 3.2 0.6 ZCR, RMSE, MFCC, ChromaSignificant values are in bold. Performance comparison of the proposed DCAT model and existing methods on the ShipEar dataset. Significant values are in bold. Classification performance on ShipEar dataset The performance of DCAT model on the ShipEar dataset is measured based on the specific class-by-class classification report on the model that includes the performance metrics of precision, recall, and F1-score, and visualization tools, including the confusion matrix, ROC curves, t-SNE plot, and Precision-Recall curves. Classification report on DCAT model shows good performance on ShipEar dataset as explained in Table 7. In the case of Class A, the model is perfectly accurate (1.0000) and nearly perfect in recall (0.9998), and the model has a perfect F1-score (1.0000), which means that the model is incredible at determining true positives, but it makes no false positives. Class B also has the highest accuracy and recall of 1.0000, giving a F1-score of 0.9997 that reveals the balanced performance of the model. Class C indicates that the precision also slightly declines to 0.9876, but the recall is 1.0000; therefore, F1-score is 0.9798. This indicates a small trade-off between precision and recall yet it has excellent classification performance. Class D also works well on precision and recalls at 1.0000, and F1-score of 1.0000 which means a perfect classification. The Class E model has a precision of 0.9896 and perfect recall (1.0000) and a solid F1-score of 0.9899, which further demonstrates its high performance. The accuracy of 0.9916 overall validates that the model is effective in all the categories with a high level of precision, recall and F1-scores in every single classification. This high performance throughout the ShipEar dataset supports the high classification performance of the model. Table 7DCAT model’s performance evaluation on objective metrics on ShipEar dataset.ClassesPrecisionRecallF1-ScoreClass A10.99981Class B110.9997Class C0.987610.9798Class D10.99981Class E0.989610.9899Accuracy0.99160.99160.9916 DCAT model’s performance evaluation on objective metrics on ShipEar dataset. Likewise, the Fig. 9 indicating the confusion matrix of the DCAT model on the ShipEar dataset demonstrates the impressive performance of the model in all classes. In the case of Class A, the model performs 100% accuracy, and there are no misclassifications. Class B also works with 100% perfection and recall. In Class C, the model has a high true publication rate of 98.9 with a low rate of misclassification of 1.1%. Class D demonstrates perfect classification with a 100% precision and recall. In Class E, the model attains 98.8% accuracy which is slightly misclassified at 1.2%. On the whole, the confusion matrix indicates that the model performs very well with limited misclassification across all classes. Confusion matrix for the DCAT model on the ShipEar dataset classification. Moreover Fig. 10 indicates the ROC curves of the proposed DCAT model on the ShipEar dataset which indicates excellent classification results. The Area under the Curve (AUC) for both the micro-average and macro-average ROC curves is 1.00, which implies that all classes are perfectly classified. Class A through Class E as well exhibit perfect performance of an AUC of 1.00. All the classes ROC curves are almost the same, and they all tend to approach the upper-left corner, which represents an ideal balance between the True and False Positive one. These findings also prove the outstanding ability of the model to classify the ShipEar data set with an absolute accuracy. ROC curve for proposed DCAT model for UTAR classification on ShipEar dataset. The Precision-Recall Curve of the proposed DCAT model on ShipEar data is provided in Fig. 11. This model demonstrates great accuracy and recalls in all classes, and every class has a perfect score of Average Precision (AP) of 1.00. The perfect AP of 1.00 confirmed in the micro-average curve proves that the overall classification performance is excellent. It is important to note that the curve with maximum F1 score has an R (Recall) value of 1.00, P (Precision) value of 0.99, and an F1-score of 0.99 demonstrating a near-perfect precision and recall balance of the model across all target classes. These findings confirm the high classification performance of DCAT model of ShipEar. Precision-recall curve for DCAT model on the ShipEar dataset. Lastly, Fig. 12 shows the t-SNE plot of the DCAT model of ShipEar dataset. The plot also indicates clearly well-separated clusters of each class, with each class being characterized by a specific color: Class A (gray), Class B (brown), Class C (blue), Class D (green), and Class E (light blue). The distinction between the classes is apparent, meaning that the model is effective in separating the classes. The t-SNE plot shows that DCAT model has learned meaningful features because the classes are spread out and do not overlap in the two-dimensional space. This also supports the fact that the model can categorize the ShipEar data with a high degree of accuracy. t-SNE visualization for the DCAT model on the ShipEar dataset for UTAR. Comparison against standard models using combined feature extraction methods Table 8 compares the proposed DCAT model with the existing models namely VGG-1955, InceptionV356, ResNet-1857 and AlexNet58 when using the same combined feature extraction methods (ZCR, RMSE, MFCC, Chroma) on the DeepShip data and applied across all the models. The comparison analyzes the models based on accuracy, inference time, FLOPs (Floating Point Operations). Table 8Comparison in terms of Inference time, Flops and parameters against existing models using features extraction such ZCR, RMSE, MFCC, Chroma for DeepShip Dataset.ModelsAccuracy %Inference Time [s]Flops [G]Params [M]VGG-1990.433020.47143.7InceptionV390.582215.0121.8ResNet-1893.141713.7511.7AlexNet94.53107.5462.3 Proposed DCAT 98.84 0.7 6.1 3.2 Significant values are in bold. Comparison in terms of Inference time, Flops and parameters against existing models using features extraction such ZCR, RMSE, MFCC, Chroma for DeepShip Dataset. Significant values are in bold. The proposed DCAT model consistently outperforms baseline models on all metrics. It has an accuracy of 98.84, which is much higher in comparison to VGG-19 (90.43%), InceptionV3 (90.58%), ResNet-18 (93.14%), and AlexNet (94.53%). The DCAT model is also highly optimized in terms of inference time, which is 0.7 s to predict, significantly exceeding inference times of the other models, which vary between 10 s (AlexNet) and 30 s (VGG-19). This makes the proposed model more efficient in real-time applications. The FLOPs and parameters of the proposed DCAT model are also very low in comparison with the standard models. The DCAT model has 6.1 G FLOPs and 3.2 million parameters, which is computationally efficient and offers a good balance between performance and resource consumption. AlexNet, by contrast, has increased FLOPs (7.54 G) and parameters (62.3 M) and VGG-19, InceptionV3, and ResNet-18 are also much more expensive in terms of computation. The experimental findings confirm the effectiveness and efficiency of the proposed DCAT model in underwater acoustic recognition tasks using the same feature extraction methods, thus it is a tool of choice in resal-world applications. Ablation study Table 9 provides an ablation study comparing the performance of the proposed DCAT model in a variety of configurations with different setups of Depthwise Convolution Blocks, the number of attention heads (H) in the Dual Transformer Branches, and Adaptive Feature Fusion. Table 9Ablation study of the proposed DCAT model considering different configurations of depthwise convolution blocks, dual transformer branches, and adaptive feature fusion.ModelDepthwiseconvolution blocksDual transformer brachesAdaptivefeature fusionSimple concatenationAccuracyBranch 1Branch 2DCATN = 8H = 0H = 0✕✓78.43N = 0H = 4H = 4✓✕79.64N = 6H = 4H = 2✓✕88.76N = 4H = 0H = 8✓✕84.28N = 0H = 4H = 0✓✕80.35N = 4H = 4H = 4✓✕93.45N = 4H = 4H = 8✕✓95.21N = 4H = 4H = 8✓✕ 98.84 Significant values are in bold. Ablation study of the proposed DCAT model considering different configurations of depthwise convolution blocks, dual transformer branches, and adaptive feature fusion. Significant values are in bold. The findings indicate that the accuracy of the model differs greatly according to the configuration. With N = 8, no Dual Transformer Branches (Both branches set to H = 0), and only 8 Depthwise Convolution Blocks the model achieves an accuracy of 78.43%. This setup is only a starting point in which the network is not utilizing the full potential of the transformer or fusion mechanisms. The performance of the model increases as we modify the configurations. As an example, with both N = 0 and both transformer branches attention head set to H = 4, the accuracy reaches 79.64% when depthwise Convolution Blocks are reduced to N = 0. Additional optimization gives rise to greater improvements. When Depthwise Convolution Blocks are configured to N = 6 and H = 4 (Branch 1) and 2 (Branch 2) the accuracy increases to 88.76%. A different arrangement, with N = 4, H = 0 of Branch 1, and H = 8 of Branch 2 gives the same accuracy of 84.28, which also gives an idea that still much can be improved. Moreover, in the cases when Branch 1 has no depthwise convolution block with N = 0 and H = 0, and Branch 2 has H = 8, the model attains an accuracy of 84.28%. The actual improvements are realized when Adaptive Feature Fusion is turned on. Using N = 4 depthwise convolution blocks and H = 4 on each of the two branches of MHA transformers and Adaptive Feature Fusion instead of simple concatenation, the accuracy increases to 93.45%. The model achieves 95.21% accuracy by adding more branch 2 transformer heads to H = 8 but with simple concatenation in place of Adaptive Feature Fusion. The optimal result is the best performance of 98.84% with depthwise convolution blocks N = 4, H = 4 block of branch 1 and H = 8 block of branch 2 of transformer and with Adaptive Feature Fusion turned on and without simple concatenation. The set up shown represents the most effective set-up to maximize the performance of DCAT model attaining the highest accuracy of 98.84 on DeepShip dataset. These findings demonstrate that the model obtains the best performance when the depthwise convolution is integrated with two transformer branches and adaptive feature fusion. The research identifies the relevance of adjusting these elements to achieve the greatest effectiveness of the model in the underwater acoustic recognition work. Conclusion This paper introduced a new Underwater Acoustic Target Recognition (UATR) model, the Depthwise Separable Convolutional Adaptive Transformer (DCAT), that can effectively combine both depthwise separable convolutions and adaptive multi-branch transformer attention functions to extract local and global features simultaneously on underwater acoustic signals. The proposed DCAT architecture integrates the advantages of convolutional efficiency with transformer-based contextual modeling to provide a powerful tradeoff between accuracy and generalization and computational efficiency, which are vital in sonar and underwater monitoring systems today. The advantages of DCAT were confirmed through experimental assessment on two benchmark datasets DeepShip and ShipsEar, with performance of 98.84% (DeepShip) and 99.16% (ShipsEar), but with an extremely lightweight design that has only 0.52 million parameters and 6.1 million FLOPs. These findings validate the ability of this model to provide near real-time inference (0.7 ms per sample) on very high-end platforms (e.g. NVIDIA RTX 2080 Ti), and can be deployed on intelligent sonar and embedded acoustic recognitions platforms, but actual results will depend on the hardware in place. In addition, the combination of Zero-Crossing Rate (ZCR), Root Mean square Energy (RMSE), Mel- Frequency Cepstral Coefficients (MFCC), and Chroma features were very important in enhancing feature discrimination and strength regarding variable underwater acoustic environments. Although these have been made, some limitations remain; the model uses predetermined feature extraction and thus does not adopt adaptability in very dynamic or noisy circumstances and may be limited to hardware throughput due to the resource constraints of embedded systems, especially in resource-limited embedded systems. The next generation of this project will be to improve robustness in extreme noise, to include self-supervised and semi-supervised learning models that remove the reliance on labeled data, and to optimize the architecture to support edge computing and real-time computing. The next upgrades to the framework should be the expansion of the framework to end-to-end learning pipelines and the combination of unsupervised feature representation, which will enhance adaptability in unpredictable acoustic settings. Overall, the proposed DCAT model defines a new performance efficient performance-efficiency point of underwater acoustic target recognition which provides a scalable, accurate, and computationally compact solution to future-generation autonomous sonar, marine surveillance, and underwater intelligent sensing systems making DCAT a potentially promising platform to build upon in the future concerning adaptable acoustic signal processing and smart maritime sensing systems. Author contributions Nahid-Al Mahmud, Tao Zhang, Yasir Iqbal: Conceptualization, Methodology, Software, Visualization, Investigation, Writing- Original draft preparation. Farhana Bari Sumona, Anjum Iqbal, Yanzhang Geng: Data curation, Validation, Supervision, Resources, Writing—Review & Editing. Wajid Khan, Basem Abu Zneid, Ievgen Zaitsev: Project administration, Supervision, Resources, Writing—Review & Editing. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Data availability The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Nahid-Al Mahmud and Yasir Iqbal contributed equally to this work. Contributor Information Tao Zhang, Email: zhangtao@tju.edu.cn. Ievgen Zaitsev, Email: zaitsev@cnpfm.kyiv.ua. References Associated Data Data Availability Statement The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.