Echofilter: A Deep Learning Segmentation Model Improves the Automation, Standardization, and Timeliness for Post-Processing Echosounder Data in Tidal Energy Streams

Scott C. Lowe, Louise P. McGarry, Jessica Douglas, Jason Newport, Sageev Oore, Christopher Whidden, Daniel J. HasselmanarXiv 2022 · DOI · OpenReview

Abstract

Understanding the abundance and distribution of fish in tidal energy streams is important to assess risks presented by introducing tidal energy devices to the habitat. However tidal current flows suitable for tidal energy are often highly turbulent, complicating the interpretation of echosounder data. The portion of the water column contaminated by returns from entrained air must be excluded from data used for biological analyses. Application of a single conventional algorithm to identify the depth-of-penetration of entrained air is insufficient for a boundary that is discontinuous, depth-dynamic, porous, and varies with tidal flow speed. Using a case study at a tidal energy demonstration site in the Bay of Fundy, we describe the development and application of a deep machine learning model with a U-Net based architecture. Our model, Echofilter, was highly responsive to the dynamic range of turbulence conditions and sensitive to the fine-scale nuances in the boundary position, producing an entrained-air boundary line with an average error of 0.33m on mobile downfacing and 0.5-1.0m on stationary upfacing data, less than half that of existing algorithmic solutions. The model's overall annotations had a high level of agreement with the human segmentation, with an intersection-over-union score of 99% for mobile downfacing recordings and 92-95% for stationary upfacing recordings. This resulted in a 50% reduction in the time required for manual edits when compared to the time required to manually edit the line placement produced by the currently available algorithms. Because of the improved initial automated placement, the implementation of the models permits an increase in the standardization and repeatability of line placement.

Introduction

Tidal energy represents a significant frontier in renewable energy, yet its deployment requires careful ecological monitoring. Researchers utilize echosounders to monitor fish populations in high-flow environments. However, the extreme turbulence characteristic of these sites introduces "entrained air"—bubbles pulled deep into the water column—which creates significant acoustic noise.

Schematic of an uplooking echosounder showing the complex boundaries in a tidal water column.

As shown in Figure 1, identifying the "observable" water requires accurately mapping the entrained-air line. This boundary is often porous and indistinct, making traditional algorithmic masking unreliable. Historically, this has necessitated labor-intensive manual cleaning by expert analysts, creating a bottleneck in data processing and introducing subjective variability.

Methods

To address these challenges, we developed Echofilter, a deep learning segmentation model. The architecture is based on a U-Net framework enhanced with EfficientNet MBConv blocks.

Several key strategies were employed to optimize performance:

  • Multi-Plane Prediction: The model is designed to predict 10 distinct planes simultaneously, including the sea surface, seafloor, and regions of "bad data" or passive recording.
  • Zoom+Repeat Strategy: To maintain high vertical resolution without overwhelming the network, the model first identifies the general region of interest in a full echogram and then performs a second, high-resolution pass on that specific area.
  • Orientation-Aware Training: The model was trained on both upfacing (bottom-mounted) and downfacing (vessel-mounted) data, allowing it to learn the unique acoustic signatures associated with different sensor perspectives.

Datasets

The model was developed using a comprehensive dataset from the Bay of Fundy, Nova Scotia, specifically from the Minas Passage and Grand Passage sites. These locations offer a range of flow speeds, from 2.5 m/s to 5 m/s.

Summary of recordings and pings used for model development and evaluation.

As detailed in Figure 2, the training set comprised millions of pings across thousands of recordings, providing a robust foundation for the model to learn the complexities of turbulent acoustic environments. Expert human annotations served as the ground truth for all evaluations.

Key Results

Echofilter significantly outperformed industry-standard algorithmic solutions (e.g., Echoview) across all primary metrics:

  • Precision: The model achieved an average boundary error of 0.33 m for mobile data and 0.5–1.0 m for stationary data.
  • Consistency: Intersection-over-Union (IoU) scores reached 99% for mobile recordings and 92–95% for stationary recordings, demonstrating high alignment with expert human judgment.
  • Operational Efficiency: Most importantly, using Echofilter as a pre-processing step reduced the time required for manual auditing by 50%. In cases with particularly poor data quality, time savings reached up to 59%.

Discussion and Conclusion

Beyond raw speed, Echofilter provides a critical benefit in standardization. By providing a repeatable, objective baseline for noise removal, it eliminates the "subjective drift" that occurs between different human analysts.

While the model shows some limitations in identifying localized "bad data" patches (such as interference from equipment), its ability to capture the complex shape of turbulence boundaries is a major advancement. Echofilter is not intended to replace human experts but to serve as a powerful tool that streamlines the path from raw data to biological insight. To support the broader research community, the model and its source code have been released under an open-source license.

Summary updated

Abstract

Understanding the abundance and distribution of fish in tidal energy streams is important to assess risks presented by introducing tidal energy devices to the habitat. However tidal current flows suitable for tidal energy are often highly turbulent, complicating the interpretation of echosounder data. The portion of the water column contaminated by returns from entrained air must be excluded from data used for biological analyses. Application of a single conventional algorithm to identify the depth-of-penetration of entrained air is insufficient for a boundary that is discontinuous, depth-dynamic, porous, and varies with tidal flow speed. Using a case study at a tidal energy demonstration site in the Bay of Fundy, we describe the development and application of a deep machine learning model with a U-Net based architecture. Our model, Echofilter, was highly responsive to the dynamic range of turbulence conditions and sensitive to the fine-scale nuances in the boundary position, producing an entrained-air boundary line with an average error of 0.33m on mobile downfacing and 0.5-1.0m on stationary upfacing data, less than half that of existing algorithmic solutions. The model's overall annotations had a high level of agreement with the human segmentation, with an intersection-over-union score of 99% for mobile downfacing recordings and 92-95% for stationary upfacing recordings. This resulted in a 50% reduction in the time required for manual edits when compared to the time required to manually edit the line placement produced by the currently available algorithms. Because of the improved initial automated placement, the implementation of the models permits an increase in the standardization and repeatability of line placement.
BibTeX
@misc{lowe2022echofilter,
  author = {Scott C. Lowe and Louise P. McGarry and Jessica Douglas and Jason Newport and Sageev Oore and Christopher Whidden and Daniel J. Hasselman},
  title = {{Echofilter: A Deep Learning Segmentation Model Improves the Automation, Standardization, and Timeliness for Post-Processing Echosounder Data in Tidal Energy Streams}},
  year = {2022},
  eprint = {2202.09648},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2202.09648},
}
Figure