Non-Contact Phenotypic Measurement and Body Mass Prediction of

To address systematic errors in prawn total length measurement caused by natural curvature and challenges in body mass prediction under random postures in aquaculture, this study proposes an automated phenotypic measurement and body mass prediction framework integrating skeleton-based nonlinear length estimation and multi-view feature analysis. Instance segmentation (YOLO11n-seg) obtains prawn head and abdomen masks. A skeleton-based procedure combining Zhang-Suen thinning, branch pruning, and graph-based path extraction obtains the projected body centerline for curved length measurement, while a curvature index quantifies body curvature. Body mass prediction models are built for side-view and top-view data, with SHAP analysis interpreting feature contributions. A unified random forest model using consistent morphological features enables body mass prediction across side-view and top-view samples. When evaluated against the independently acquired manual two-segment reference, the skeleton-based method achieved an MAE of 0.399 cm for severely curved prawns, representing reductions of 70.3%, 66.0%, and 21.9% relative to the straight-line method, MBR method, and Zhang-Suen baseline, respectively. On an independent test set, side-view and top-view models achieved mean absolute percentage errors of 5.73% and 5.82%, respectively. The unified RF model achieved an overall MAPE of 5.54%, and paired comparisons did not detect statistically significant differences from the corresponding viewpoint-specific models on either test subset. These results demonstrate the effectiveness of the proposed framework for non-contact prawn phenotyping under controlled imaging conditions.