Norsk English

Ensemble Learning for Robust Fish Species Identification Across Diverse Underwater Environments

The goal of this thesis is to explore and implement ensemble learning techniques to create a robust system for identifying fish species across different underwater environments, using existing models trained for specific conditions.

Contact persons

Fish species identification in underwater environments. Ill.: SINTEF

Description

While individual models may perform well in their respective environments (e.g., different cameras, water clarity, lighting conditions, or geographical locations), there is often a need for a unified system capable of adapting to the variations encountered in real-world scenarios.

This research will focus on combining multiple pre-trained models through ensemble learning methods such as bagging, boosting, stacking, and voting. The thesis will explore the following aspects:

  • Model Aggregation: Combine models that have been trained on fish species detection in different environments (e.g., clear vs. murky waters, different types of underwater cameras, shallow vs. deep water).
  • Performance Evaluation: Test and validate ensemble methods to see how they improve robustness and accuracy compared to individual models, particularly in unseen or challenging environments.
  • Weighted Voting: Implement techniques to weigh models differently based on environmental factors (e.g., camera type, water quality) or model confidence, to optimize species identification.
  • Generalization: Ensure the ensemble model generalizes well across various conditions, reducing the risk of poor performance in unfamiliar or mixed environments.

The project will include dataset preparation from various underwater sources, model training and integration, and performance benchmarking. This research has potential applications in fisheries monitoring, biodiversity studies, and marine conservation.

The project can be further expanded to include the recognition of plankton (both phytoplankton and zooplankton) in collaboration with SINTEF Ocean, utilizing equipment such as SilCam - SINTEF

Key Deliverables

  • A detailed study on the use of ensemble learning for improving fish species identification across varying conditions.
  • A comparative analysis of different ensemble methods.
  • A functional ensemble model and evaluation using real-world underwater data.

References

  1. Frederik E.T. Schöller, Martin K. Plenge-Feidenhans’l, Jonathan D. Stets, Mogens Blanke. Object Detection at Sea Using Ensemble Methods Across Spectral Ranges, IFAC-PapersOnLine, Volume 54, Issue 16, 2021, Pages 1-6, ISSN 2405-8963
  2. Arthur Vilhelm, Matthieu Limbert, Clément Audebert, and Tugdual Ceillier. Ensemble Learning Techniques for Object Detection in High-Resolution SatelliteImages. 2022. arXiv:2202.10554. 
  3. Model ensembling
  4. DTO-BioFlow

Fish Species Detection in the Oslofjord Using Underwater Camera Technology

A fish species detection model was applied to a video stream from the Oslofjord, sourced from the Live Streaming OsloMet Oceanlab ANERIS project, featuring a low-cost underwater camera observatory prototype (available at https://www.youtube.com/watch?v=RLsfsCDvd9E&ab_channel=OsloMetOceanlab%2FHavlaboratoriet). The object detection model was developed and integrated during the Iliad Hackathon 2024, "Digital Twins of the Ocean," organized by the Iliad Project on September 11-12, 2024, by the Oslofjord Digital Twin team.

 

A fish species detection model was applied to a video stream from the Oslofjord, sourced from the Live Streaming OsloMet Oceanlab ANERIS project, featuring a low-cost underwater camera observatory prototype (available at https://www.youtube.com/watch?v=RLsfsCDvd9E&ab_channel=OsloMetOceanlab%2FHavlaboratoriet). The object detection model was developed and integrated during the Iliad Hackathon 2024, "Digital Twins of the Ocean," organized by the Iliad Project on September 11-12, 2024, by the Oslofjord Digital Twin team.