A Survey of Underwater Acoustic Data Classification Methods Using Deep Learning for Shoreline Surveillance

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Cohen's Kappa coefficient is a useful metric for imbalanced datasets, but the interpretation of its value is problematic.

“Cohen’s Kappa is an useful metric for imbalanced datasets; however, the interpretation of its value is problematic [84].”

• cites [84] Delgado, R., and Tibau, X.A. (2019). Why Cohen’s Kappa should be avoided as performance measure in classification. PLoS ONE, 14. sourceDisputed

  The Introduction says Kappa can be inadequate when class distributions are imbalanced, and the Conclusion says it cannot serve as a dependable classification measure in such settings. The paper does discuss difficulties interpreting Kappa, but it contradicts the claim that Kappa is useful for imbalanced datasets. (LLM, conf=0.89)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

May combine two claims.

The auditory-inspired CNN had a high accuracy (87.2%), outperforming a standard CNN (83.2%), a stack of autoencoders (81.4%), wavelet transform with PCA (74.6%), and a support-vector machine (68.2%).

“The results showed that the proposed auditory-inspired CNN had a high accuracy (87.2%), outperforming a standard CNN (83.2%), a stack of autoencoders [113] (81.4%) and various traditional object recognition algorithms, such as a combination of wavelet transform with PCA [114] (74.6%) and a support-vector machine [115] (68.2%).”

• cites [101] Shen, S., Yang, H., Yao, X., Li, J., Xu, G., and Sheng, M. (2020). Ship type classification by convolutional neural networks with auditory-like mechanisms. Sensors, 20. sourcePartly supported

  Section 4.2 (page unavailable) says the proposed model outperformed manually designed features and improved on a previous auditory-inspired CNN. Our copy does not substantiate the five quoted accuracy figures or the claimed ranking against a standard CNN, stacked autoencoders, wavelet transform with PCA, and an SVM. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [102] Shen, S., Yang, H., and Li, J. (2019). Improved Auditory Inspired Convolutional Neural Networks for Ship Type Classification. OCEANS 2019-Marseille, IEEE. sourceNot checked — no full text of the source
• cites [103] Shen, S., Yang, H., Li, J., Xu, G., and Sheng, M. (2018). Auditory Inspired Convolutional Neural Networks for Ship Type Classification with Raw Hydrophone Data. Entropy, 20. sourceDisputed

  Section 5.2 (page unavailable) reports 79.2% accuracy for the auditory-inspired CNN and 73.2% for the spectral-feature CNN, contrary to the claimed 87.2% and 83.2%. The source supports the general finding that the proposed model outperformed its CNN baseline, but the other named methods and their claimed accuracies are not reported in the cited passage. (LLM, conf=0.90)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured
• cites [113] Cao, X., Zhang, X., Yu, Y., and Niu, L. (2016, January 16–18). Deep learning-based recognition of underwater target. Proceedings of the 2016 IEEE International Conference on Digital Signal Processing (DSP), Beijing, China. sourceNot checked — no full text of the source
• cites [114] Shi, M., and Xu, X. (2013, January 5–8). Underwater target recognition based on wavelet packet entropy and probabilistic neural network. Proceedings of the 2013 IEEE International Conference on Signal Processing, Communication and Computing (ICSPCC 2013), Kunming, China. sourceNot checked — no full text of the source

May combine two claims.

Shen et al. do not report the exact devices used, the precise period of data collection, or the labelled data used, making it impossible to reproduce the reported results even though the original raw data are freely available.

“However, Shen et al. [101] do not report the exact devices that were used, the precise period of data collection, or the labelled data used. Therefore, even though the original raw data are freely available, it is impossible to reproduce the results reported in that work.”

• cites [101] Shen, S., Yang, H., Yao, X., Li, J., Xu, G., and Sheng, M. (2020). Ship type classification by convolutional neural networks with auditory-like mechanisms. Sensors, 20. sourceDisputed

  Section 4.1 (Experimental Dataset; page unavailable) directly contradicts the claim that Shen et al. did not report the exact device: it identifies an Ocean Sonics icListen AF hydrophone. It gives a 29-month collection duration and ship-type categories, but not precise calendar dates or the labels for individual recordings. Our copy does not establish that the raw data are freely available or that reproduction is impossible. (LLM, conf=0.85)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

The SSD model extends VGG-16 with six additional convolutional layers, each producing feature maps at the scale of the targets.

“The SSD model extends VGG-16 with six additional convolutional layers, each producing feature maps at the scale of the targets.”

• cites [64] Liu, 2016, SSD: Single shot multibox detector, European Conference on Computer Vision sourcePartly supported

  The Model section (p. 3) says SSD uses VGG-16 as its base and adds convolutional feature layers for predictions at multiple scales. The PASCAL VOC2007 section (p. 7) names six layers used for prediction, but two—conv4_3 and conv7—come from the base network. It does not support calling all six additional layers. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In SSD, additional layers detect and classify large objects using low-resolution feature maps, while high-resolution maps are used to detect and classify smaller objects.

“The additional layers defining an SSD model detect and classify large objects using feature maps of low resolutions, whereby high-resolution maps are used to detect and classify smaller objects.”

• cites [64] Liu, 2016, SSD: Single shot multibox detector, European Conference on Computer Vision sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Interference of the direct path and the surface reflected path creates the Lloyd mirror pattern, which makes object detection and classification in shallow waters challenging.

“The interference of these two paths creates the so-called Lloyd mirror pattern [24], which makes object detection and classification in shallow waters (such as in a harbour or port) a challenging task [25]”

• cites [24] McMahon, 2012, Horizontal Lloyd mirror patterns from straight and curved nonlinear internal waves, J. Acoust. Soc. Am. sourceNot checked — no full text of the source
• cites [25] Ferguson, E.L., Ramakrishnan, R., Williams, S.B., and Jin, C.T. (2017, January 5–9). Convolutional neural networks for passive monitoring of a shallow water environment using a single sensor. Proceedings of the 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), New Orleans, LA, USA. sourcePartly supported

  The Introduction (p. 1) describes clutter, noise, and multipath reflection in shallow waters such as harbours and ports, and says conventional time-delay estimation is limited when the Lloyd’s mirror effect is not discernible. This supports the general difficulty of acoustic monitoring there, but the paper links the effect specifically to limitations in ranging, not to difficulty in both object detection and classification. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

The Single-Shot Detector (SSD) attempts to balance high accuracy and performance on object detection based on a feed-forward convolutional network.

“Another neural network that fits into the one-stage detector category is the Single-Shot Detector (SDD) [64], which attempts to balance high accuracy and performance on object detection based on a feed-forward convolutional network.”

• cites [64] Liu, 2016, SSD: Single shot multibox detector, European Conference on Computer Vision sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Faster R-CNN eliminates the selective search method used in R-CNN and Fast R-CNN and uses a separate CNN (Region Proposal Network) to predict region proposals directly from convolutional feature maps.

“The Faster R-CNN algorithm [71] eliminates the selective search method used in both R-CNN and Fast R-CNN, and uses a separate CNN (called Region Proposal Network (RPN)) to predict region proposals directly from the convolutional feature maps.”

• cites [71] Ren, 2016, Faster R-CNN: Towards real-time object detection with region proposal networks, IEEE Trans. Pattern Anal. Mach. Intell. sourceSupported

May combine two claims.

Traditional ML classification metrics propagate underlying biases, ignore the trustworthiness of negative instances, and do not take into account chance-level performance.

“These traditional metrics, however, propagate the underlying biases, ignore the trustworthiness of negative instances and do not take into account the chance-level performance [79,81].”

• cites [79] Powers, 2011, Evaluation: From Precision, Recall and F-Factor to ROC, Informedness, Markedness & Correlation, Int. J. Mach. Learn. Technol. sourceNot checked yet

  Not checked in the free run.

• cites [81] Valverde-Albacete, F.J., and Peláez-Moreno, C. (2014). 100% classification accuracy considered harmful: The normalized information transfer factor explains the accuracy paradox. PLoS ONE, 9. sourcePartly supported

  The Introduction (page unavailable) explains that accuracy can favor a majority-class classifier on imbalanced data, and Results > Two performance measures based on perplexity discusses the accuracy of a clueless classifier. These passages support concerns about class bias and chance performance, but the paper does not establish that traditional metrics generally ignore the trustworthiness of negative instances. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

MCC is a more reliable metric, assigning a high score only if the prediction was correct for the four base categories, taking into account the size of P and N.

“MCC is a more reliable metric, whereby a high score is assigned only if the prediction was correct for the four base categories (TP, FP, TN and FP), taking into account the size of P and N [85].”

• cites [85] Chicco, D., and Jurman, G. (2020). The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation. BMC Genom., 21. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

AUSCC was introduced to provide an unbiased evaluation of ML models for classification of underwater acoustic signals, taking into account sample sizes per class.

“The notion of area under the sample complexity curve (AUSCC) was recently introduced in [88] to provide an unbiased evaluation of ML models for the classification of underwater acoustic signals, taking into account the sample sizes per class.”

• cites [88] Valdenegro-Toro, M., Preciado-Grijalva, A., and Wehbe, B. (2021). Pre-trained Models for Sonar Images. arXiv. sourcePartly supported

  In “Transfer Learning on Watertank Dataset” (p. 5), the paper introduces AUSCC to compare model accuracy across different numbers of training samples per class. It evaluates classification of sonar images, but does not describe AUSCC as an unbiased evaluation or establish its use for underwater acoustic signals generally. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

The method in [120] requires the fundamental frequencies of the target object to be known beforehand, and assumes they are independent of environmental conditions.

“However, the fundamental frequencies of the target object need to be known beforehand, and it is assumed that they are independent of environmental conditions (which may not be true in the general case).”

• cites [120] Bach, N.H., Vu, L.H., and Nguyen, V.D. (2021). Classification of Surface Vehicle Propeller Cavitation Noise Using Spectrogram Processing in Combination with Convolution Neural Network. Sensors, 21. sourcePartly supported

  Section 2.3 says the methods require a target frequency window to be defined and that unsuitable input parameters can make detection unfeasible. It does not say the target’s fundamental frequencies must be known beforehand. Nor does it assume independence from environmental conditions; Section 2.2 says underwater signals are influenced by many environmental factors. No page numbers are available in our copy. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

May combine two claims. The quote only partly states this claim.

A system using convolutional filters as feature extractors, two-dimensional convolutional layers, and a Bidirectional-LSTM with a trainable attention module achieved high-accuracy results when tested on a shallow-water dataset from the Indian Ocean.

“This system was tested on a dataset obtained in shallow waters in the Indian Ocean, achieving high-accuracy results.”

• cites [125] Kamal, 2021, Passive sonar automated target classifier for shallow waters using end-to-end learnable deep convolutional LSTMs, Eng. Sci. Technol. Int. J. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Results obtained with the method in [106] suggest that not using a CNN to learn the features saves computing time, as no redundant features are generated.

“Results obtained with this method suggest that by not using a CNN to learn the features saves computing time, as no redundant features are generated.”

• cites [106] Wang, X., Liu, A., Zhang, Y., and Xue, F. (2019). Underwater Acoustic Target Recognition: A Combination of Multi-Dimensional Fusion Features and Modified Deep Neural Network. Remote Sens., 11. sourcePartly supported

  The Introduction says that directly using a DNN can generate redundant features and waste computing resources. Section 3 says the paper uses a GMM to reduce redundant features and the size of its modified DNN. This supports the resource-saving rationale, but the source describes a DNN rather than a CNN and does not report a measured computing-time saving. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

May combine two claims.

The combination of cepstral filtering and CNNs greatly extends detection ranges and robustness to noise compared to an algorithmic method, but has slightly reduced accuracy at short range due to expressive bias.

“The combination of cepstral filtering and CNNs greatly extends detection ranges and robustness to noise compared to an algorithmic method, but has slightly reduced accuracy at short range due to expressive bias [25].”

• cites [25] Ferguson, E.L., Ramakrishnan, R., Williams, S.B., and Jin, C.T. (2017, January 5–9). Convolutional neural networks for passive monitoring of a shallow water environment using a single sensor. Proceedings of the 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), New Orleans, LA, USA. sourcePartly supported

  The Comparison of Range and Detection Methods (p. 4) reports that CNNs estimate vessel range beyond 180 m, where the algorithmic method fails, but have greater error and significant bias below 180 m. The Conclusions (p. 4) describe robustness to changes in SNR. Neither passage attributes the short-range bias to an expressive limitation, and the paper calls the bias significant rather than slight. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

The combination of ACT with the EFP method achieved higher accuracy with less resources (memory and time) compared to Random Forest and Support Vector Machines.

“The combination of ACT with the EFP method was compared against Random Forest and Support Vector Machines, achieving higher accuracy with less resources (memory and time).”

• cites [117] Miao, 2021, Underwater Acoustic Signal Classification Based on Sparse Time–Frequency Representation and Deep Learning, IEEE J. Ocean. Eng. sourceSupported

May combine two claims.

All algorithms in [6] presented excellent performance, with an error rate below 5%.

“All algorithms presented excellent performance, with an error rate below 5%.”

• cites [6] Choi, J., Choo, Y., and Lee, K. (2019). Acoustic classification of surface and underwater vessels in the ocean using supervised machine learning. Sensors, 19. sourceSupported

May combine two claims.

Testing with real-world images revealed poor performance of AlexNet overall, whereas GoogLeNet obtained an accuracy of 91.6% when trained with input data with various noise levels.

“The accuracy obtained testing the CNN with the real-world set of images revealed a poor performance of AlexNet overall, whereas GoogLeNet obtained an accuracy of 91.6% when trained with input data with various noise levels.”

• cites [130] Nguyen, H.T., Lee, E.H., and Lee, S. (2019). Study on the Classification Performance of Underwater Sonar Image Classification Based on Convolutional Neural Networks for Detecting a Submerged Human Body. Sensors, 20. sourceSupported

May combine two claims.

A data-augmentation technique using an RBM autoencoder to reconstruct original signals produced mean squared error between reconstruction and original samples in the range 0.0115 to 0.5649, meaning the difference was minor and the properties were maintained.

“The mean squared error (MSE) between the reconstruction and the original samples was in the range of 0.0115 to 0.5649, meaning that the difference between the two signals was minor and the properties were maintained.”

• cites [123] Luo, 2021, An Underwater Acoustic Target Recognition Method Based on Combined Feature With Automatic Coding and Reconstruction, IEEE Access sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

A model trained with synthetic data generated by neural style transfer can improve both classification accuracy and convergence time.

“The data were used to balance an uneven dataset, verifying that a model trained with synthetic data can improve both the classification accuracy and the convergence time.”

• cites [132] Ge, Q., Ruan, F., Qiao, B., Zhang, Q., Zuo, X., and Dang, L. (2021). Side-Scan Sonar Image Classification Based on Style Transfer and Pre-Trained Convolutional Neural Networks. Electronics, 10. sourceSupported

May combine two claims.

CNNs, even when trained with distinct datasets from the target, are capable of learning relevant features, improving classification accuracy using transfer learning.

“Results show that CNNs (even when trained with distinct datasets from the target) are capable of learning relevant features, improving the accuracy of classification using transfer learning.”

• cites [131] Valdenegro-Toro, M. (2017, January 19–22). Best practices in convolutional networks for forward-looking sonar image recognition. Proceedings of the OCEANS 2017-Aberdeen, Aberdeen, UK. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

A real-time automatic target recognition method combining a transformer attention mechanism and the YOLOv5s model, named TR-YOLOv5s, was applied to side-scan sonar images.

“Side-scan sonar images were used as inputs to a real-time automatic target recognition (ATR) method that applied a combination of a transformer attention mechanism [138] and the YOLOv5s model [133], named TR–YOLOv5s.”

• cites [138] Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R. (2017). Attention is All you Need. Advances in Neural Information Processing Systems, Curran Associates, Inc.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [133] Yu, Y., Zhao, J., Gong, Q., Huang, C., Zheng, G., and Ma, J. (2021). Real-Time Underwater Maritime Object Detection in Side-Scan Sonar Images Based on Transformer-YOLOv5. Remote Sens., 13. sourceSupported

May combine two claims.

Use of augmented data in the training stage generalised well to real-world problems, providing a classification accuracy of 91.6% using GoogLeNet as the classifier.

“The use of augmented data in the training stage proved to generalise well to the real-world problems, providing a classification accuracy of 91.6% using GoogLeNet as the classifier.”

• cites [130] Nguyen, H.T., Lee, E.H., and Lee, S. (2019). Study on the Classification Performance of Underwater Sonar Image Classification Based on Convolutional Neural Networks for Detecting a Submerged Human Body. Sensors, 20. sourceSupported

May combine two claims.

In an experiment using synthetic data from the KRAKEN normal-mode propagation model, the convolutional model produced the best accuracy results but showed high variability with respect to input data representation, whereas the feed-forward neural network was not affected.

“The convolutional model produced the best accuracy results; however, it also showed high variability with respect to the input data representation (where a phone-space representation presented as a more suitable input type, as discussed in Section 3), whereas the FNN was not affected.”

• cites [6] Choi, J., Choo, Y., and Lee, K. (2019). Acoustic classification of surface and underwater vessels in the ocean using supervised machine learning. Sensors, 19. sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

Much work in underwater acoustic data classification is conducted using synthetic data only or on a limited set of real data augmented with semisynthetic examples for training.

“Therefore, much work in this area is conducted using synthetic data only [124], or on a limited set of real data augmented with semisynthetic examples for training [62,128,139].”

• cites [124] Denos, K., Ravaut, M., Fagette, A., and Lim, H.S. (2017). Deep learning applied to underwater mine warfare. OCEANS 2017-Aberdeen, IEEE. sourceNot checked — no full text of the source
• cites [62] Le, 2020, Deep gabor neural network for automatic detection of mine-like objects in sonar imagery, IEEE Access sourceSupported
• cites [128] Huo, 2020, Underwater object classification in sidescan sonar images using deep transfer learning and semisynthetic training data, IEEE Access sourceSupported
• cites [139] Berg, H., and Hjelmervik, K.T. (2018, January 22–25). Classification of anti-submarine warfare sonar targets using a deep neural network. Proceedings of the OCEANS 2018 MTS/IEEE Charleston, Charleston, SC, USA. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Accuracy is used for comparative analysis in various publications in this area.

“although accuracy is used for comparative analysis in the various publications in this area”

Classified as probably not a substantial claim.

Ignoring sonar sensing variables limits the use of machine-learning methods in applications that demand a high level of trust, such as defence systems or inspection of underwater built structures.

“Ignoring these variables limits the use of ML methods in applications that demand a high-level of trust, such as in defence systems or in the inspection of underwater built structures (such as deep-sea mining facilities)”

Probably not a substantial claim: may not pass S3 (load-bearing).

There is some evidence that certain cells in the mammalian visual cortex can be approximated by Gabor filters.

“There is some evidence that certain cells in the mammalian visual cortex (responsible for the perception of texture) can be approximated by such filters [40]”

• cites [40] Jones, 1987, An evaluation of the two-dimensional Gabor filter model of simple receptive fields in cat striate cortex, J. Neurophysiol. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Fast R-CNN is faster than R-CNN because the feature map is generated only once, rather than for every selected region.

“Fast R-CNN is faster than R-CNN since in the former the feature map is generated only once, rather than for every region selected (as imposed by the latter).”

• cites [70] Girshick, R. (2015, January 7–13). Fast R-CNN. Proceedings of the IEEE International Conference on Computer Vision, Santiago, Chile. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The characteristics of sound propagation underwater are largely ignored in most literature concerning machine-learning-based underwater acoustic signal classifiers, even though they are essential for interpreting the accuracy of classification results.

“the characteristics of sound propagation underwater, however, are largely ignored in most literature concerning machine-learning-based underwater acoustic signal classifiers, even though it is essential for interpreting the accuracy of the classification results”

May combine two claims.

Most existing research on the application of machine-learning methods for the classification of underwater acoustic data ignores sonar sensing variables and does not evaluate classification tasks with respect to sonar system performance in distinct oceanographic situations.

“most existing research on the application of machine-learning methods for the classification of underwater acoustic data (as we shall see below) ignores these effects, and does not evaluate the classification tasks with respect to the performance of the sonar system in distinct oceanographic situations”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Scattering due to rough boundaries or small obstacles causes loss in the acoustic signal, and scattering happens with wavelengths at the order of the obstacles, causing parts of the acoustic field to be randomised.

“Scattering, due to rough boundaries or small obstacles, is another process that causes loss in the acoustic signal. In contrast to reflection, scattering happens with wavelengths at the order of the obstacles, causing parts of the acoustic field to be randomised [26]”

• cites [26] Weill, A. (2014). Acoustic Waves, Scattering. Encyclopedia of Remote Sensing, Springer. sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

Spectrographic and cepstographic representations of acoustic data are analogous to visual images in that they represent a signal across two dimensions.

“Spectrographic (and cepstographic) representations of acoustic data are analogous to visual images in that they represent a signal across two dimensions (x, y spatial dimensions for images and time, frequency for spectrograms)”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

VGGnets seem to be the model of choice in most work on the classification of objects from sonar data, as they present a good trade-off between classification accuracy and hyperparameter complexity.

“VGGnets seem to be the model of choice in most work on the classification of objects from sonar data, as they present a good trade-off between classification accuracy and hyperparameter complexity”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Scattering due to rough surfaces causes frequency-dependent attenuation of the acoustic field, whereas volume scattering decreases with depth and varies at distinct periods of the day.

“Scattering due to rough surfaces cause a frequency-dependent attenuation of the acoustic field, whereas volume scattering (usually due to near-surface bubbles, bubble clouds or biological obstacles, such as air-filled swim bladders and zooplankton) decreases with depth and presents variations at distinct periods of the day [20]”

• cites [20] Kuperman, W., and Roux, P. (2007). Underwater Acoustics. Springer Handbook of Acoustics, Springer. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

A negative sound-speed gradient causes sound rays to bend downwards, a positive gradient causes the opposite, and sound bends toward regions of low sound speed.

“Therefore, a negative sound–speed gradient (e.g., in a thermocline) causes sound rays to bend downwards; the opposite happens with a positive gradient (e.g., in deep ocean where changes in pressure are greater than changes in temperature). In other words, sound bends toward regions of low sound speed [20]”

• cites [20] Kuperman, W., and Roux, P. (2007). Underwater Acoustics. Springer Handbook of Acoustics, Springer. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The open source availability of large, annotated image datasets has provided the appropriate basis for the rapid development of image classification.

“The open source availability of large, annotated image datasets has provided the appropriate basis for the rapid development of this field [52]”

• cites [52] LeCun, 2015, Deep learning, Nature sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Some recent methods in acoustic processing use Gabor-filter banks on sound spectra as a preprocessing step and achieve accurate results on acoustic processing tasks such as classification of environmental sounds, music genre recognition, and speech analysis.

“some recent methods in acoustic processing use Gabor-filter banks on the sound spectra as a preprocessing step, achieving accurate results on acoustic processing tasks, such as the classification of environmental sounds [41], music genre recognition [42], and speech analysis [43,44], to cite a few”

• cites [41] Souli, S., and Lachiri, Z. (2012, January 21–25). Environmental sound classification using log-Gabor filter. Proceedings of the International Conference on Signal Processing Proceedings, ICSP, Beijing, China. sourceNot checked — no full text of the source
• cites [42] Costa, Y., Oliveira, L., Koerich, A., and Gouyon, F. (2013). Music genre recognition using gabor filters and LPQ texture descriptors. Iberoamerican Congress on Pattern Recognition, Springer. sourceNot checked — no full text of the source
• cites [43] Ezzat, T., Bouvrie, J.V., and Poggio, T.A. (2007, January 27–31). Spectro-temporal analysis of speech using 2-D Gabor filters. Proceedings of the Interspeech 8th Annual Conference of the International Speech Communication Association, Antwerp, Belgium. sourceNot checked — no full text of the source
• cites [44] He, L., Lech, M., Maddage, N., and Allen, N. (2009, January 10–12). Stress and emotion recognition using log-Gabor filter analysis of speech spectrograms. Proceedings of the 2009 3rd International Conference on Affective Computing and Intelligent Interaction and Workshops, Amsterdam, The Netherlands. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Misclassifications are likely to result if the confounding effects of temperature, depth and boundary conditions are not taken into account by the system.

“misclassifications are likely to result if the confounding effects of temperature, depth and boundary conditions are not taken into account by the system”

Classified as probably not a substantial claim.

Various data-augmentation techniques are described in the literature to cope with the scarcity of annotated sonar datasets from real maritime missions.

“various data-augmentation techniques are described in the literature (summarised in Section 4.1)”

• AssumptionRests on an assumption: Obtaining sonar data from real maritime missions is complex.

Probably not a substantial claim: may not pass S3 (load-bearing).

The influences of oceanographic properties on sound propagation paths can be classified into three basic classes: very short range, deep water, and shallow water propagation.

“In more general terms, the influences of oceanographic properties on the sound propagation paths can be classified into three basic classes: very short range, deep water, and shallow water propagation [20]”

• cites [20] Kuperman, W., and Roux, P. (2007). Underwater Acoustics. Springer Handbook of Acoustics, Springer. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Time-frequency analysis does not perform any classification or categorisation of the information within the data; it only changes how the data are represented and presented.

“However, these approaches are nothing more than changing the way the data are represented and presented. They do not perform any sort of classification or categorisation of the information within the data”

Probably not a substantial claim: may not pass S3 (load-bearing).

The SSD model is based on the VGG convolutional neural network.

“The model is based on the VGG convolutional neural network [57].”

• cites [57] Simonyan, K., and Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Automatic classification of vessel acoustic signals is challenging because the signal depends on vessel speed, the age and state of the propulsion system, highly variable background noise, and the diversity of sound propagation mechanisms in the ocean.

“The automatic classification of this type of signal is a challenging task, as the signal is also dependent on the vessel’s speed, the age and state of the propulsion system, the highly variable background noise and the diversity of sound propagation mechanisms in the ocean”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Sound propagation in the ocean depends on properties of the water column (temperature, salinity and pressure) and on ocean floor effects such as scattering and reverberation.

“Sound propagation in the ocean is dependent on properties of the water column (temperature, salinity and pressure) and effects related to the ocean floor (scattering and reverberation)”

• cites [20] Kuperman, W., and Roux, P. (2007). Underwater Acoustics. Springer Handbook of Acoustics, Springer. sourceNot checked — no full text of the source
• cites [21] Abraham, D.A. (2019). Underwater Acoustic Signal Processing Modeling, Detection, and Estimation, Springer Nature. sourceNot checked — no full text of the source
• cites [22] Urick, R. (1979). Sound Propagation in the Sea. sourceNot checked — no full text of the source

May combine two claims.

There is a clear scarcity of publicly available, labelled datasets of underwater acoustic signals, preventing reproducibility and repeatability of results and greatly hindering the development of the field.

“There is a clear scarcity of publicly available, labelled datasets of underwater acoustic signals, preventing reproducibility and repeatability of the results, greatly hindering the development of this field”

• AssumptionRests on an assumption: Obtaining sonar data from real maritime missions is complex.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The acoustic signal produced by a vessel moving in the sea is mainly composed of a broadband component generated by the propeller and its hydrodynamic interactions, and a narrow band component due to the propulsion system and other mechanical parts.

“The acoustic signal produced by a vessel moving in the sea is mainly composed of a broadband component (with a continuous spectrum), that is generated by the propeller and its hydrodynamic interactions; and a narrow band component (whose spectrum consists of line components at discrete frequencies), owing to the propulsion system and other mechanical parts [19]”

• cites [19] Vahidpour, 2015, An automated approach to passive sonar classification using binary image features, J. Mar. Sci. Appl. sourceNot checked — no full text of the source

May combine two claims.

In high-trust applications, classic acoustic processing strategies are still the first-choice methods because these effects can be explicitly accounted for.

“In these applications, classic acoustic processing strategies (not involving machine-learning algorithms) are still the first-choice methods because these effects can be explicitly accounted for”

Probably not a substantial claim: may not pass S3 (load-bearing).

These invariance properties allowed the definition of a taxonomy matching appropriate metrics with qualities of datasets such as representativeness of class distribution, reliability of class labels, and unimodality or multimodality of classes.

“These properties allowed the definition of a taxonomy matching the appropriate metrics with qualities of the datasets, such as representativeness of class distribution, reliability of class labels, and the unimodality or multimodality of the classes.”

• cites [80] Sokolova, 2009, A systematic analysis of performance measures for classification tasks, Inf. Process. Manag. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

An autoencoder attempts to replicate the components of the input that represent common patterns or signals of interest while ignoring noise.

“Briefly, an AE attempts to replicate the components of the input that represent the common patterns, or signals of interest, while ignoring noise.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Most applications of CNNs to acoustic signals are based on adapting methods initially developed for images.

“Therefore, most applications of CNNs to acoustic signals are based on adapting the methods initially developed for images.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Autoencoders can be used as a preprocessing step before a classical classifier or as components within a deep CNN.

“Since they can remove noise and simplify data, AEs can be used as a preprocessing step before a classical classifier (such as a SVM) [75] or even as components within a deep CNN [76,77].”

• cites [75] Ju, Y., Guo, J., and Liu, S. (2015, January 17–19). A Deep Learning Method Combined Sparse Autoencoder with SVM. Proceedings of the 2015 International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery, Xi’an, China. sourceNot checked — no full text of the source
• cites [76] Sun, 2017, Generalized extreme learning machine autoencoder and a new deep neural network, Neurocomputing sourceNot checked — no full text of the source
• cites [77] Huang, 2020, A deep learning algorithm using a fully connected sparse autoencoder neural network for landslide susceptibility prediction, Landslides sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

UATC-DenseNet was designed to classify underwater acoustic targets from raw audio signals.

“Similarly, a CNN model called Underwater Acoustic Target Classification DenseNet (UATC-DenseNet) [92] was designed to classify underwater acoustic targets from raw audio signals.”

• cites [92] Doan, 2020, Underwater acoustic target classification based on dense convolutional neural network, IEEE Geosci. Remote Sens. Lett. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

GANs are two-component deep neural networks that may or may not have convolutional layers.

“Generative Adversarial Networks (GANs) [78] are two component deep neural networks that may or may not have convolutional layers.”

• cites [78] Goodfellow, 2020, Generative Adversarial Networks, Commun. ACM sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Powers argued that informedness and markedness overcome issues related to most common metrics such as accuracy, precision, and recall.

“Recently, Powers [79] argued that the notions of informedness and markedness overcome the issues related to most common metrics, such as accuracy, precision, recall and other related concepts.”

• cites [79] Powers, 2011, Evaluation: From Precision, Recall and F-Factor to ROC, Informedness, Markedness & Correlation, Int. J. Mach. Learn. Technol.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

GAN components are trained simultaneously in a two-player adversarial zero-sum game.

“The two components making up the GANs architecture are trained simultaneously in a two-player adversarial zero-sum game.”

Probably not a substantial claim: may not pass S3 (load-bearing).

GAN training stops when the discriminator classifies most synthetic examples as real, meaning the generator is reliably creating data indistinguishable from the training set.

“The training of a GAN stops when the discriminator classifies as real most of the synthetic examples generated by the generator, meaning that the first component is reliably creating data that are indistinguishable from the training set.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

UATC-DenseNet is composed of a deep CNN with dense architecture trained to recognise 12 classes of signals.

“This model is composed of a deep CNN with dense architecture that was trained to recognise 12 classes of signals.”

• cites [92] Doan, 2020, Underwater acoustic target classification based on dense convolutional neural network, IEEE Geosci. Remote Sens. Lett. sourceNot checked — no full text of the source

May combine two claims.

These findings have yet to be applied for the evaluation of underwater acoustic automatic classification tasks, whose current literature assumes accuracy or mAP as the main evaluation metric.

“These findings have yet to be applied for the evaluation of underwater acoustic automatic classification tasks, whose current literature (as we shall see later in this paper) assumes accuracy or mAP as the main evaluation metric.”

Probably not a substantial claim: may not pass S3 (load-bearing).

GANs have recently been used as a data-augmentation strategy for generating new datasets for training underwater acoustic object classifiers.

“Recently, GANs have been used as an important data-augmentation strategy, generating new datasets for training underwater acoustic object classifiers, as we shall see further in this paper.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The GAN generator is trained to reproduce patterns observed in input data, generating synthetic data points that could have been part of the input set.

“The first component, called generator, is trained to reproduce the patterns observed in the input data, generating as output synthetic data points that could have been part of the input (training) set.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Autoencoders are an example of a neural network model that abstracts out desired patterns from noisy data.

“Autoencoders (AE) are an example of a NN model that executes this task [72,73].”

• cites [72] Vincent, 2010, Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion, J. Mach. Learn. Res. sourceNot checked — no full text of the source
• cites [73] Ng, 2011, Sparse autoencoder, CS294A Lect. NotesNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The major contribution of R-CNN, Fast R-CNN, and Faster R-CNN is the improvement of accuracy and performance combined.

“The major contribution of these methods is the improvement of the accuracy and performance combined.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Autoencoders are more akin to a nonlinear denoising filter than a classifier because they do not attempt to categorise components within the input data.

“In this way, AEs are more akin to a nonlinear denoising filter than a classifier since they do not attempt to categorise components within the input data [74].”

• cites [74] Wang, W., Huang, Y., Wang, Y., and Wang, L. (2014, January 23–28). Generalized Autoencoder: A Neural Network Framework for Dimensionality Reduction. Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition Workshops, Columbus, OH, USA. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Autoencoders are an example of self-supervised learning because they do not need labelled input data to train from.

“Furthermore, AEs are an example of self-supervised learning (even though they are trained by means of supervised learning methods) since they do not need labelled input data to train from as they merely try to replicate a lower dimensionality/complexity version of the input.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The dataset used for UATC-DenseNet was labelled by a sonar expert, containing 11 different classes of sounds and one class corresponding to ambient sound.

“The dataset used was labelled by a sonar expert, containing 11 different classes of sounds and one class corresponding to the ambient sound.”

• cites [92] Doan, 2020, Underwater acoustic target classification based on dense convolutional neural network, IEEE Geosci. Remote Sens. Lett. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Although AUC has been used to classify the accuracy of target detection algorithms, its extension to multiclass classification is currently not totally understood.

“Although AUC has been used to classify the accuracy of target detection algorithms [86], its extension to the analysis of multiclass classification is currently not totally understood [80,87].”

• cites [86] Uzair, 2021, Bio-Inspired Video Enhancement for Small Moving Target Detection, IEEE Trans. Image Process. sourceNot checked — no full text of the source
• cites [80] Sokolova, 2009, A systematic analysis of performance measures for classification tasks, Inf. Process. Manag. sourceNot checked — no full text of the source
• cites [87] Lachiche, N., and Flach, P. (2003, January 21–24). Improving Accuracy and Cost of Two-Class and Multi-Class Probabilistic Classifiers Using ROC Curves. Proceedings of the ICML’03, Washington, DC, USA.Not checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

A system combining an unsupervised SOM network with LVQ was proposed to provide high performance in classifying the roughness of the ocean floor from raw sonar data.

“In this work, a system that combines an unsupervised self-organised mapping (SOM) network [99] with learning vector quantisation (LVQ) [100] was proposed to provide high performance in classifying the roughness of the ocean floor from raw sonar data.”

• cites [98] Chakraborty, B., Kodagali, V., Baracho, J., and Joseph, A. (2002). System for Classifying Seafloor Roughness. (6763303B2), U.S. Patent.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [99] Kohonen, 1982, Self-organized formation of topologically correct feature maps, Biol. Cybern. sourceNot checked — no full text of the source
• cites [100] Touretzky, D., Mozer, M.C., and Hasselmo, M. (1996). Generalized Learning Vector Quantization. Advances in Neural Information Processing Systems, MIT Press.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

The GAN discriminator is trained to determine if the input is from the real dataset or synthetic data from the generator.

“The other component, called discriminator, is trained to determine if the input is from the real dataset or synthetic data from the first GAN component.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Sokolova and Lapalme present an analysis of 24 evaluation metrics for ML classification tasks, showing invariance properties across the metrics.

“Sokolova and Lapalme [80] present an analysis of 24 evaluation metrics for ML classification tasks, showing the existence of invariance properties across the metrics.”

• cites [80] Sokolova, 2009, A systematic analysis of performance measures for classification tasks, Inf. Process. Manag. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

In general, the use of bioinspired preprocessing steps in ML classification procedures has been shown to reduce the complexity of ML training, accelerating learning while avoiding local maxima.

“Nevertheless, in general, the use of bioinspired preprocessing steps in ML classification procedure has been shown to reduce the complexity of the ML training procedure, accelerating the learning process while avoiding local maxima.”

May combine two claims.

Conventional algorithmic approaches to vessel classification and ranging from hydrophone data perform well at close range (<180 m) but their performance degrades quickly at longer ranges.

“Conventional algorithmic approaches to vessel classification and ranging from hydrophone data perform well at close range (<180 m) but their performance degrades quickly at longer ranges.”

• cites [121] Ferguson, 2005, Sensor position estimation and source ranging in a shallow water environment, IEEE J. Ocean. Eng. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based).

Underwater acoustic data for training machine-learning classification algorithms are scarce.

“The scarcity of underwater acoustic data for training ML classification algorithms is a common statement in most related publications.”

May combine two claims.

The use of wavelet transform was shown to be effective at enhancing noise robustness, achieving 99.7% classification accuracy with an SNR of 0 dB, greatly outperforming six other CNN architectures.

“The use of wavelet transform was shown to be effective at enhancing noise robustness in the system, achieving 99.7% classification accuracy with an SNR of 0 dB, greatly outperforming six other CNN architectures.”

• cites [118] Kim, 2021, A method for underwater acoustic signal classification using convolutional neural network combined with discrete wavelet transform, Int. J. Wavelets Multiresolut. Inf. Process. sourceNot checked — no full text of the source

May combine two claims.

Prefiltering in the time-frequency domain further improves classification, as does the inclusion of convolutional layers within the classifier model.

“Moreover, prefiltering in the time–frequency domain further improves classification as does the inclusion of convolutional layers within the classifier model.”

May combine two claims.

The results presented in [105] were only compared with other metaheuristic algorithms training the same neural network, but not against state-of-the-art deep-learning methods.

“The results presented were only compared with other metaheuristic algorithms training the same neural network, but not against state-of-the-art deep-learning methods.”

• cites [105] Khishe, 2019, Passive sonar target classification using multi-layer perceptron trained by salp swarm algorithm, Ocean Eng. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Finding suitable hyperparameters for deep-learning architectures remains a challenging task.

“However, a challenging task still remains, that is, to find suitable hyperparameters for the DL architectures.”

May combine two claims.

An experiment with three different SNR (without noise; 5 dB and 10 dB) compared the accuracy of the proposed model with a fractal-based method, showing that NN have better noise robustness for all analysed scenarios.

“An experiment with three different SNR (without noise; 5 dB and 10 dB) compared the accuracy of the proposed model with a fractal-based method, showing that NN have a better noise robustness for all analysed scenarios.”

• cites [19] Vahidpour, 2015, An automated approach to passive sonar classification using binary image features, J. Mar. Sci. Appl. sourceNot checked — no full text of the source

May combine two claims.

The data used in [105] are not available to reassess the results, making reproduction or comparison with modern classification systems impossible.

“The data used are not available to reassess these results, making reproduction, or even comparison with the performance of modern classification systems, impossible.”

• cites [105] Khishe, 2019, Passive sonar target classification using multi-layer perceptron trained by salp swarm algorithm, Ocean Eng. sourceNot checked — no full text of the source

May combine two claims.

Transfer learning has advantages in machine-learning algorithm development, such as shorter training times and enhanced robustness from using models that have already passed initial tests.

“TL has many advantages in the development of ML algorithms, such as shorter training times and the enhanced robustness of using models that have already passed initial tests.”

• cites [127] Pan, 2009, A survey on transfer learning, IEEE Trans. Knowl. Data Eng. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

YOLOv5s reused weights trained on the COCO dataset, improving generalisation ability and reducing final training time.

“The YOLOv5s model reused the weights trained on the COCO dataset, improving the generalisation ability and reducing the final training time.”

Classified as probably not a substantial claim. May combine two claims.

Confidentiality issues, the cost and complexity of running real maritime missions, and the expensive task of curating, annotating and storing large amounts of data are the main reasons for the nondisclosure of datasets.

“Confidentiality issues, the cost and complexity of running real maritime missions, combined with the expensive task of curating, annotating and storing large amounts of data, are the main reasons for the nondisclosure of datasets.”

Classified as probably not a substantial claim.

CNNs are data hungry, requiring a large amount of labelled data to properly train and validate models.

“CNNs have proved to be an important tool for automatic data classification. However, they are data hungry, as a large amount of labelled data points is needed to properly train and validate the models.”

May combine two claims.

A typical GAN applied to generate sonar images of mine-like objects produced only a moderate improvement in accuracy of a CNN+HGP classifier (1.36 percentage points, maximum accuracy 81.56%) and a similar improvement for a stand-alone CNN (56.5% to 57.2%).

“The results show only a moderate improvement in the accuracy of the CNN+HGP classifier when using the augmented dataset (1.36 percentage points, achieving a maximum accuracy of 81.56%); a similar improvement was observed with respect to a stand-alone CNN (an improvement from 56.5% without the data augmentation to 57.2% with the GAN-generated dataset).”

• cites [147] Phung, S.L., Nguyen, T.N.A., Le, H.T., Chapple, P.B., Ritz, C.H., Bouzerdoum, A., and Tran, L.C. (2019, January 2–4). Mine-Like Object Sensing in Sonar Imagery with a Compact Deep Learning Architecture for Scarce Data. Proceedings of the 2019 Digital Image Computing: Techniques and Applications (DICTA), Perth, Australia. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

There is uncertainty about the realism of the augmented dataset used for the Deep Gabor Neural Network.

“However, there is uncertainty about the realism of the augmented dataset used.”

May combine two claims.

Use of the augmented dataset in [118] reduced overfitting and improved generalisation ability.

“The results reported showed a reduction in overfitting and an improvement in generalisation ability with the use of the augmented dataset.”

• cites [118] Kim, 2021, A method for underwater acoustic signal classification using convolutional neural network combined with discrete wavelet transform, Int. J. Wavelets Multiresolut. Inf. Process. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

A data-augmentation procedure convolving an acoustic signal and an impulse response signal, then adding white Gaussian noise, generated a set of 30,000 images of 30 classes of underwater acoustic signals.

“Analogously, a data-augmentation procedure was performed in [118] by convolving an acoustic signal and an impulse response signal. Then, white Gaussian noise was added to finally produce the extended signals. After the data-augmentation process, a set containing 30,000 images of 30 different classes of underwater acoustic signals was generated.”

• cites [118] Kim, 2021, A method for underwater acoustic signal classification using convolutional neural network combined with discrete wavelet transform, Int. J. Wavelets Multiresolut. Inf. Process. sourceNot checked — no full text of the source

May combine two claims.

Without a common body of data, comparing and benchmarking solutions is virtually impossible, rendering this type of research largely insubstantial.

“However, it also renders this type of research largely insubstantial, since without a common body of data, comparing and benchmarking solutions is a virtually impossible task.”

May combine two claims.

No architecture showed superior performance over all training/test scenarios, known as the no free-lunch theorem.

“Moreover, although various machine-learning models were trained, there was no architecture that showed a superior performance over all training/test scenarios, which is a common issue found in machine learning, known informally as the “no free-lunch theorem” [136], where no one model is the best under all situations.”

• cites [136] Dowe, 2011, No Free Lunch versus Occam’s Razor in Supervised Learning, Proceedings of the Algorithmic Probability and Friends. Bayesian Prediction and Artificial Intelligence—Papers from the Ray Solomonoff 85th Memorial ConferenceNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Transfer learning is applied as an alternative to the scarcity of real-world data to train deep-learning models, enabling reuse of well-established results obtained from other domains.

“Transfer learning is applied as an alternative to the scarcity of real-word data to train DL models, enabling the reuse of well-established results obtained from other domains.”

May combine two claims.

Most approaches do not assess the characteristics of generated data in comparison with the original ones, but measure quality through classification accuracy, which represents a problem in the generalisation of such techniques.

“Although there has been an undeniable advance in this area, most approaches do not assess the characteristics of the data generated in comparison with the original ones, but measure the quality of the procedure through the accuracy of the classification, which represents a problem in the generalisation of such techniques.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Data augmentation has proven to be a viable solution for the lack of annotated datasets, becoming an initial step for most studies related to the underwater acoustic domain.

“Data augmentation has proven itself to be a viable solution for the lack of annotated datasets, becoming an initial step for most of the studies related to the underwater acoustic domain.”

May combine two claims.

The presence of sonar echo generated from the mine itself can mask the real shape of the mine-like object, but it had a minor influence on learning the shadow concept.

“The presence of sonar echo generated from the mine itself can mask the real shape of the MLO, but it had a minor influence on learning the shadow concept [124].”

• cites [124] Denos, K., Ravaut, M., Fagette, A., and Lim, H.S. (2017). Deep learning applied to underwater mine warfare. OCEANS 2017-Aberdeen, IEEE. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Recent literature has described important positive results with the use of transfer learning techniques to tackle the scarcity of annotated datasets.

“Recent literature has described important positive results with the use of transfer learning (TL) techniques to tackle this issue as well (cf Section 3).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

To the best of the authors' knowledge, no reference evaluates the impact of environmental parameters on classifiers' performance.

“To the best of our knowledge, there is no reference that evaluates the impact of these parameters with classifiers’ performance.”

• AssumptionRests on an assumption: Environmental factors such as time of day, season of the year, geographic regions, type of sensor, depth, and pressure influence the quality of underwater acoustic data.