A domain-adapted contrastive learning framework for label-free fish feeding related activity discrimination using ensemble detection and trajectory-based behavioral features

Fish feeding behavior is a vital predictor of fish activity and well-being in aquaculture, but manual assessment is time consuming, subjective, and hard to generalize. In this study, we propose a label-free framework for fish feeding-related activity discrimination by integrating ensemble object detection, multi-object tracking, trajectory-based feature engineering and domain-adapted contrastive representation learning. Fish videos from the Fish4Knowledge dataset were used for experimentation, comprising 18 video sequences and approximately 8,427 extracted frames. YOLOv8n and YOLOv8s pre-trained on DeepFish were used as fish detectors, while their predictions were fused using Weighted Box Fusion (WBF). The tracking of detected fish was done by the ByteTrack system in order to keep the continuity of the trajectory. Five behavior features were identified based on valid fish trajectory including mean distance, standard deviation of distance, direction entropy, surface occupancy, and distance between fish to define the intensity of movement, variability of movement, randomness of direction, vertical location, and social behavior, respectively. The descriptors were then encoded using a SimCLR-style domain adaptive contrastive pretext task by creating positive and negative samples based on the temporal relationships and cross fish relationships. The representations were further clustered using K-means to find the unique states of activities. Among 847 valid trajectories, the proposed technique gave three clusters, including 788, 40, and 19 of trajectories respectively, with a Silhouette Score of 0.6547. The Silhouette Score for the raw features was equal to 0.4733 while for generic contrastive baseline it was equal to 0.2728. These findings prove that domain-adapted contrastive representation learning enhances the separability of trajectory-based fish activity patterns without the need for labeled behavioral information.

Authors

Institutions

Publication Details

Journal
Frontiers in Sustainable Food Systems
Published
2026-11-18
DOI
https://doi.org/10.3389/fsufs.2026.1984206
Primary Topic
Water Quality Monitoring Technologies
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A domain-adapted contrastive learning framework for label-free fish feeding related activity discrimination using ensemble detection and trajectory-based behavioral features

Ganesh Kumar Chellamani, Aishwarya N., Komaleeswari P.
3 links
Frontiers in Sustainable Food Systems
Water Quality Monitoring Technologies
article

A domain-adapted contrastive learning framework for label-free fish feeding related activity discrimination using ensemble detection and trajectory-based behavioral features

Ganesh Kumar Chellamani, Aishwarya N., Komaleeswari P.
article en

Abstract

Fish feeding behavior is a vital predictor of fish activity and well-being in aquaculture, but manual assessment is time consuming, subjective, and hard to generalize. In this study, we propose a label-free framework for fish feeding-related activity discrimination by integrating ensemble object detection, multi-object tracking, trajectory-based feature engineering and domain-adapted contrastive representation learning. Fish videos from the Fish4Knowledge dataset were used for experimentation, comprising 18 video sequences and approximately 8,427 extracted frames. YOLOv8n and YOLOv8s pre-trained on DeepFish were used as fish detectors, while their predictions were fused using Weighted Box Fusion (WBF). The tracking of detected fish was done by the ByteTrack system in order to keep the continuity of the trajectory. Five behavior features were identified based on valid fish trajectory including mean distance, standard deviation of distance, direction entropy, surface occupancy, and distance between fish to define the intensity of movement, variability of movement, randomness of direction, vertical location, and social behavior, respectively. The descriptors were then encoded using a SimCLR-style domain adaptive contrastive pretext task by creating positive and negative samples based on the temporal relationships and cross fish relationships. The representations were further clustered using K-means to find the unique states of activities. Among 847 valid trajectories, the proposed technique gave three clusters, including 788, 40, and 19 of trajectories respectively, with a Silhouette Score of 0.6547. The Silhouette Score for the raw features was equal to 0.4733 while for generic contrastive baseline it was equal to 0.2728. These findings prove that domain-adapted contrastive representation learning enhances the separability of trajectory-based fish activity patterns without the need for labeled behavioral information.

Frontiers in Sustainable Food SystemsVol. 10
Amrita Vishwa Vidyapeetham (IN)
Water Quality Monitoring Technologies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon