Otolith-based age determination for Baltic cod
A research project on machine learning–supported age determination of Baltic cod from otolith microscopy images. The work focuses on data consolidation, quality assessment, and the development of supervised models to support expert-based age reading.
Context
Age determination is a central task in fisheries science and stock assessment. For Baltic cod, age is commonly estimated by manually reading growth patterns in otolith cross-sections under a microscope. This process is time-consuming, relies on a small number of highly trained expert readers, and is sensitive to image quality and subjective judgement.
At the same time, the number of experienced readers is decreasing, while consistent age information remains critical for long-term monitoring and management. This raises the question of whether machine learning–based methods can support — and eventually help scale — otolith-based age determination from microscopy images.
Project focus
This project investigates how machine learning models can support age determination from otolith cross-section images of Baltic cod. The current focus lies on building the methodological and data-related foundations required for reliable model development.
Concretely, the project explores:
- supervised age prediction from microscopy images
- early quality assessment of otolith cuts and images (e.g. suitability for reading)
Collaboration
This project is carried out as a small, closely coordinated collaboration:
- Uwe Krumme — domain expert in otolith reading and fisheries science; initiator of the project and primary source of methodological and biological expertise
- Farnaz Kharratahari — student research assistant (HiWi), contributing to data preparation, exploratory analysis, and model experimentation
- Jan Meischner — research coordination and methodological development
My role
I coordinate the overall project design and methodological direction. My responsibilities include:
- structuring the research questions and project phases
- designing the data pipeline and evaluation strategy
- supporting model selection, training, and interpretation
- integrating domain expertise with data-driven methods
Implementation
Current and planned implementation work includes:
- ingestion and handling of high-resolution microscopy images in proprietary formats
- systematic preprocessing and annotation of otolith images
- consolidation and validation of associated metadata collected over multiple decades
- exploratory development of machine learning models for age prediction
- formulation of image and cut quality classification tasks as a prerequisite for reliable age estimation
At the current stage, data preparation and quality assessment constitute a substantial part of the work.
Stack
Outcomes
- A consolidated and documented dataset of otolith cross-section images and associated metadata
- A clearer understanding of data quality constraints and their implications for model development
- Baseline approaches for age prediction and image quality assessment
- A reproducible experimental setup to support further development and evaluation
Learnings
- Preparing a usable dataset from nearly two decades of manually recorded data is a significant challenge in itself.
- Inconsistencies and irregularities in metadata require careful validation and explicit assumptions.
- Data quality and documentation are not peripheral concerns but central determinants of what forms of modeling are feasible.