Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cross-Domain Knowledge Transfer for Underwater Acoustic Classification Using Pre-trained Models

Amirmohammad Mohammadi, Tejashri Kelhe, Davelle Carreiro, Alexandra Van Dine, Joshua Peeples

TL;DR

The paper investigates cross-domain transfer learning for underwater acoustic classification by comparing AudioSet-pretrained PANNs with ImageNet-pretrained TIMMs on the DeepShip passive sonar dataset. Using a consistent preprocessing and augmentation pipeline, it demonstrates that ImageNet-pretrained TIMMs can surpass audio-pretrained models, particularly at various sampling rates, and that data resolution interacts with pre-training to influence performance. Through extensive experiments and interpretability analyses (Grad-CAM), the work highlights the practical potential of cross-domain transfer learning to address data scarcity in UATR and suggests directions for incorporating self-supervised and multi-modal approaches. The findings have implications for deploying efficient, robust underwater classifiers in real-world maritime applications where labeled data are limited.

Abstract

Transfer learning is commonly employed to leverage large, pre-trained models and perform fine-tuning for downstream tasks. The most prevalent pre-trained models are initially trained using ImageNet. However, their ability to generalize can vary across different data modalities. This study compares pre-trained Audio Neural Networks (PANNs) and ImageNet pre-trained models within the context of underwater acoustic target recognition (UATR). It was observed that the ImageNet pre-trained models slightly out-perform pre-trained audio models in passive sonar classification. We also analyzed the impact of audio sampling rates for model pre-training and fine-tuning. This study contributes to transfer learning applications of UATR, illustrating the potential of pre-trained models to address limitations caused by scarce, labeled data in the UATR domain.

Cross-Domain Knowledge Transfer for Underwater Acoustic Classification Using Pre-trained Models

TL;DR

The paper investigates cross-domain transfer learning for underwater acoustic classification by comparing AudioSet-pretrained PANNs with ImageNet-pretrained TIMMs on the DeepShip passive sonar dataset. Using a consistent preprocessing and augmentation pipeline, it demonstrates that ImageNet-pretrained TIMMs can surpass audio-pretrained models, particularly at various sampling rates, and that data resolution interacts with pre-training to influence performance. Through extensive experiments and interpretability analyses (Grad-CAM), the work highlights the practical potential of cross-domain transfer learning to address data scarcity in UATR and suggests directions for incorporating self-supervised and multi-modal approaches. The findings have implications for deploying efficient, robust underwater classifiers in real-world maritime applications where labeled data are limited.

Abstract

Transfer learning is commonly employed to leverage large, pre-trained models and perform fine-tuning for downstream tasks. The most prevalent pre-trained models are initially trained using ImageNet. However, their ability to generalize can vary across different data modalities. This study compares pre-trained Audio Neural Networks (PANNs) and ImageNet pre-trained models within the context of underwater acoustic target recognition (UATR). It was observed that the ImageNet pre-trained models slightly out-perform pre-trained audio models in passive sonar classification. We also analyzed the impact of audio sampling rates for model pre-training and fine-tuning. This study contributes to transfer learning applications of UATR, illustrating the potential of pre-trained models to address limitations caused by scarce, labeled data in the UATR domain.
Paper Structure (11 sections, 1 equation, 4 figures, 2 tables)

This paper contains 11 sections, 1 equation, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Model Preparation
  4. Data Preparation
  5. Data Augmentation
  6. Experimental Procedure
  7. Results and Discussion
  8. Impact of Sampling Rate
  9. Impact of Pre-training Data
  10. Best PANN and TIMM Models Comparison
  11. Conclusion

Figures (4)

  • Figure 1: The overall framework is shown. (a) Data Preprocessing: The audio waveform is transformed into a logarithmic mel-frequency spectrogram. (b) Data Augmentation: SpectAugmentation park2019specaugment and Mixup zhang2018mixup are added during training to improve model robustness. (c) Data Analysis: Input spectrograms are processed using networks pre-trained on AudioSet (PANN kong2020panns) or ImageNet (TIMM rw2019timm).
  • Figure 2: Average test accuracy across three experimental runs at different sampling rates with $\pm1$ standard deviation for CNN14 models. Each color represents a different model, with the symbol of each model centered on the average test accuracy and the error bars show $\pm1$ standard deviation.
  • Figure 3: Average confusion matrices for best PANN model (a) CNN14-32k (70.6 $\pm$ 0.8) and best TIMM model (b) ConvNeXtV2-tiny (73.7 $\pm$ 0.8) across three experimental runs. The average test accuracy $\pm1$ standard deviation is shown in parentheses.
  • Figure 4: Visualization of Grad-CAM results for (a) CNN14 correctly classified samples, (b) CNN14 misclassified samples, (c) ConvNeXtV2-tiny correctly classified samples, and (d) ConvNeXtV2-tiny misclassified samples.