Species identification in sonar data using simulation-informed modeling

Context

Fisheries surveys rely heavily on active acoustics to estimate fish abundance and distribution. While sonar data provide high temporal and spatial coverage, species identification still depends largely on net-based sampling. These catches are sparse in time and space, invasive, and introduce uncertainty when extrapolated to continuous acoustic observations.

At the same time, fish distribution is not arbitrary. Vertical positioning and aggregation behavior are shaped by environmental conditions such as temperature, salinity, and oxygen, as well as species-specific preferences and avoidance ranges. This raises the question of whether environmental information and process-based assumptions can be used to better constrain the interpretation of sonar observations.

Project focus

This project investigates how simulation-based representations of vertical fish distribution can support the interpretation and classification of sonar data, with a primary focus on sprat and, where applicable, herring.

The current work focuses on:

• simulating vertical fish distribution under given physical conditions derived from CTD data
• comparing simulation outputs with observations from fisheries surveys, including acoustic data and depth-resolved catch information
• assessing how well simulated distributions align with observed sonar backscatter patterns

Building on this foundation, the longer-term perspective of the project is the development of a machine learning model that integrates sonar data with environmental parameters and simulation-informed priors to support species classification in acoustic observations.

The overarching motivation is to reduce reliance on lethal sampling while improving the consistency and resolution of abundance estimates derived from acoustic surveys.

Collaboration

The project is carried out in close collaboration with the Thünen Institute of Baltic Sea Fisheries (Rostock) and the Chair of Marine Data Science at the University of Rostock.

Key collaborators include:

• Stefanie Haase (Thünen Institute of Baltic Sea Fisheries)
• Stefan Lüdtke (Chair of Marine Data Science, University of Rostock)

The scientific direction and applied relevance of the project are strongly driven by the fisheries research context of the Thünen Institute, with methodological development embedded within this framework.

My role

I develop the simulation framework for vertical fish distribution and design the data-driven workflows used to compare simulated results with fisheries survey observations. This includes structuring the data pipeline, integrating environmental and acoustic data sources, and exploring how process-based assumptions can inform downstream modeling.

The project forms a core part of my doctoral research, with a focus on bridging fisheries science, environmental data, and machine learning–based analysis.

Stack

Expected outcomes

• A simulation framework for vertical fish distribution under varying environmental conditions
• A reproducible workflow linking CTD data, sonar observations, and catch information
• Quantitative comparison of simulated distributions and observed acoustic patterns
• Conceptual groundwork for machine learning models that integrate sonar data with environmental and simulation-based priors
• A contribution to methods aimed at reducing invasive sampling while improving abundance estimation in fisheries surveys