Abstract
Monitoring marine ecosystems relies heavily on acoustic sonar data, traditionally analyzed through labor-intensive manual annotation of echograms. This paper presents a deep learning framework utilizing Mask R-CNN for instance segmentation of fish schools (herring and juvenile salmon). By generating pixel-level masks rather than coarse bounding boxes, the proposed method significantly improves detection accuracy and localization while achieving a 10x increase in processing speed compared to existing object detection baselines.
Introduction
Marine biologists utilize echosounders to produce echograms—2D representations of acoustic backscatter where the vertical axis represents depth and the horizontal axis represents time. Identifying specific fish schools within these data is critical for biomass estimation and fishery management. However, manual "painting" of these schools is subjective and time-consuming. Previous automated attempts using object detection (e.g., YOLOv2) often struggle with the irregular shapes of fish schools, overlapping instances, and environmental noise, frequently resulting in overestimation of school area or fragmented detections.
Methods
The core of the proposed system is the Mask R-CNN architecture, which performs simultaneous object detection and instance segmentation. This allows the model to delineate the exact boundaries of fish schools at a pixel level.
Key methodological contributions include:
- End-to-End Processing: Unlike previous methods that required breaking echograms into small, overlapping tiles, the proposed method processes larger sections directly, reducing computational overhead.
- Transfer Learning: Models were pre-trained on the COCO dataset and fine-tuned on acoustic data.
- Backbone Optimization: Various architectures (ResNet-50, ResNet-101, ResNeXt-101) were evaluated to optimize the trade-off between inference speed and precision.
- Physical Units: Raw data is converted to Volume Backscattering Strength ($S_v$) to ensure physical consistency across different sonar settings.
Dataset: The PLHS Dataset
The study introduces the Pixel-Level Herring and Salmon (PLHS) dataset, containing 107 expert-annotated echograms from the coast of British Columbia. It includes 405 instances (153 herring, 252 juvenile salmon) with manual pixel-level labels. This dataset addresses the limitations of previous box-labeled datasets by providing ground truth for the actual morphology of fish schools.
Key Results and Discussion
The proposed Mask R-CNN framework outperformed the YOLOv2 baseline across all metrics.
- Accuracy: The system achieved a Mean Average Precision (mAP) of 92.12% at IoU=0.5, a substantial improvement over the 55.67% achieved by the previous state-of-the-art.
- Biomass Estimation: Because bounding boxes include significant "empty" water around irregular schools, they often overestimate school area by 100% or more. Pixel-precise masks provide a much more reliable basis for biomass calculation.
- Species Challenges: Juvenile salmon proved more difficult to detect than herring due to their smaller, sparser school structures and proximity to surface turbulence.
- Inference Speed: The end-to-end approach allows for processing one hour of sonar data in approximately 0.4 seconds on a mid-range GPU, representing a 10x speedup over tiling-based methods.
Conclusion
Transitioning from object detection to instance segmentation represents a major step forward in automated fisheries acoustics. The ability to precisely delineate fish schools in real-time enables more accurate biological assessments and reduces the burden on human experts. Future work will focus on expanding the species catalog and improving model robustness across different sonar hardware platforms.
Summary updated