Improving fishing ground estimation with weak supervision and meta-learning.

Area of Science:

• Marine Biology
• Data Science
• Artificial Intelligence

Background:

• Accurate fishing ground estimation is crucial for the fishing industry.
• Traditional methods often struggle with limited or imprecise data.
• Sea surface temperature patterns offer potential for predicting fishing locations.

Purpose of the Study:

• To develop a robust method for estimating fishing grounds using deep learning.
• To address the challenge of limited annotated catch data for model training.
• To leverage readily available trajectory data for improved model performance.

Main Methods:

• Utilized a deep learning-based keypoint detector for pattern recognition in sea surface temperature data.
• Implemented a training strategy combining weak supervision and meta-learning.
• Employed pre-training with trajectory data and fine-tuning with catch data, incorporating a meta-learner to mitigate label noise.

Main Results:

• The proposed method significantly improved fishing ground estimation accuracy.
• Achieved a 64% increase in F1-score compared to a baseline model using only catch data.
• Demonstrated the effectiveness of combining weak supervision and meta-learning for handling data limitations.

Conclusions:

• The integration of weak supervision and meta-learning offers a powerful solution for data-scarce scenarios in fishing ground estimation.
• Deep learning models can effectively utilize sea surface temperature patterns and trajectory data for improved predictions.
• This approach enhances the efficiency and accuracy of marine resource management.