In aquaculture production, manual or fixed-schedule feeding often fails to match the real-time feeding level of fish schools, and overfeeding can lead to feed wastage and water-quality deterioration, which has become a major bottleneck for both large-scale farming efficiency and environmental sustainability. During feeding, intense competition and jumping behaviors generate splashes of varying magnitudes, which can serve as an indirect visual proxy for hunger intensity. In this study, we constructed a frame-level splash-annotated dataset and performed data preprocessing. Building upon YOLO11 pretrained weights, we introduced a P2–P5 four-scale detection head to enhance small-splash recognition, injected EGMA into the backbone C3k2 blocks, and replaced stride-2 downsampling convolutions with a three-branch ADown operator. On the validation set, the proposed YOLO11-PEGA achieved a precision of 0.86 and a recall of 0.80, with mAP@0.5 exceeding 0.80 and mAP@0.5–0.95 exceeding 0.30. Compared with the baseline model, the parameter count was reduced by 72.3%. The results demonstrate that the proposed model maintains stable detection and evaluation performance under complex environmental conditions, providing actionable decision support for feeding-threshold setting, feeding-time determination, and feed-amount adjustment.

Large yellow croaker (Larimichthys crocea) is an economically important marine fish, yet rapid and non-destructive discrimination between sea-caught and farmed collection-source groups remains challenging. This study developed a fish-level analytical workflow integrating YOLO11n-seg instance segmentation with visible–near-infrared hyperspectral imaging to extract relative reflectance ratios from six anatomical regions. The analytical cohort comprised 258 unique fish, including 171 sea-caught and 87 farmed individuals. Source categories were assigned according to the original capture or cage-culture collection channels. Fish identity was used as the grouping unit in five-fold internal cross-validation, and classification performance was evaluated based on out-of-fold fish-level predictions. Logistic-regression models achieved AUC values ranging from 0.985 to 1.000 across the six individual anatomical regions. The combined six-region model achieved an accuracy of 0.996 (95% CI, 0.988–1.000) and an AUC of 1.000 (95% CI, 1.000–1.000) within the present cohort. For anatomical-region segmentation, the 20-image validation set contained 140 annotated instances. Bounding-box precision, recall, mAP@0.5, and mAP@0.5–0.95 were 0.976, 0.983, 0.981, and 0.700, respectively, while the corresponding mask metrics were 0.969, 0.976, 0.972, and 0.669. These results indicate that region-specific hyperspectral information can support highly accurate internal discrimination between the two collection-source groups and provide an interpretable basis for characterizing source-associated spectral differences. However, the source labels were not independently verified, potentially influential covariates were incompletely recorded, and no independent external cohort was available. Therefore, the present findings should be interpreted as internally validated collection-source discrimination rather than verified provenance authentication or evidence of external generalizability.