Deep learning model for detecting fish foraging movement and distance

Koi ( Cyprinus carpio ) is a high-value ornamental fish species with both ecological and economic importance, particularly in East Asia, Notably, few studies have conducted quantitative analysis of koi foraging behavior in intelligent aquaculture systems. The present study developed a fish group behavior observation framework combining high-precision object detection with multi-object tracking to capture fine-scale spatial and temporal dynamics of koi foraging. The framework involves (i) foraging activity tracking, with Fish Group Average Movement (FGAM) and Fish Group Central Point Tracking (FGCPT) models used to quantify activity intensity and group-level movement patterns, and (ii) a foraging behavior overview, with Fish Group Dispersion (FGD) and Fish Group Right-Side Foraging Ratio (FGRSFR) models used to assess spatial cohesion and feeding zone engagement. YOLOv8 was used for fish detection, and DeepSORT was used for identity tracking. The framework achieved 99 % detection and 85 % tracking accuracy, effectively identifying foraging phases such as aggregation and dispersal. By integrating computer vision and behavioral models, this framework provides real-time indicators to optimize feeding strategies and contributes a scalable, noninvasive solution to ecological informatics. • Novel model uses detection (YOLOv8) and tracking (Deep SORT) to analyze koi foraging. • Four models across two domains analyze fish activity and foraging behavior overview quantitatively. • High accuracy (99 % detection, 85 % tracking) achieved even with overlapping fish in complex scenes. • Night feeding showed greater activity; the model supports adaptive and sustainable aquaculture management.