PA-SWE: a posture-aware sow weight estimation framework via skeleton-calibrated multimodal fusion
Computers and Electronics in Agriculture · 2026-09 · 252 · 112075
DOI: 10.1016/j.compag.2026.112075 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computers and Electronics in Agriculture, 2026-09. Verfasst von Junbin Liu und 3 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben. Gefördert von: Science and Technology Planning Project of Guangdong Province, Guangdong Provincial Applied Science and Technology Research and Development Program, National Institute of Science and Technology for Translational Research on Health and Environment in the Amazon Region.
Abstract (Originalsprache)
Accurate body weight monitoring of gestating sows is critical for precision feeding and reproductive performance optimization. While deep learning has shown promise in image-based weight estimation, existing RGB-based methods are sensitive to posture variation, and depth-based alternatives incur higher deployment costs and environmental constraints. This study introduces Posture-Aware Sow Weight Estimation (PA-SWE), a posture‑aware RGB–skeletal framework that uses posture as an auxiliary source of structural information rather than treating it as nuisance variation. The framework comprises an upstream module that simultaneously outputs instance segmentation masks and anatomical keypoints from the top‑view RGB images, and a downstream module that fuses RGB appearance features with a skeletal graph via compact bilinear pooling to jointly predict body weight and posture. Evaluation on a commercial farm dataset containing 48 gestating sows of a single breed follows an event‑level protocol: image‑level predictions within each weighing event are aggregated using the median to produce the final event-level estimate. PA‑SWE achieves an event‑level mean absolute error of 2.814 kg, a root mean square error of 3.513 kg, and an R 2 of 0.975, together with a macro‑averaged F1‑score of 92.89% for three‑class posture classification. Pig‑level statistical analyses further showed that the error reduction relative to RGB‑only baselines is significant and consistent across individuals. These results suggest that explicit integration of skeletal cues can reduce posture‑related estimation errors and improve event‑level weight estimation in the current farm‑specific setting. The proposed framework therefore offers a practical option for contactless monitoring of gestating sows without requiring strict standing-posture standardization under crate-based monitoring conditions.
Förderung
- Science and Technology Planning Project of Guangdong Province
- Guangdong Provincial Applied Science and Technology Research and Development Program
- National Institute of Science and Technology for Translational Research on Health and Environment in the Amazon Region
Zitieren
APA
Liu, J., Xiao, D., Tan, Z., & Huang, Y. (2026). PA-SWE: a posture-aware sow weight estimation framework via skeleton-calibrated multimodal fusion. Computers and Electronics in Agriculture, 252, 112075. https://doi.org/10.1016/j.compag.2026.112075
Vancouver
Liu J, Xiao D, Tan Z, Huang Y. PA-SWE: a posture-aware sow weight estimation framework via skeleton-calibrated multimodal fusion. Computers and Electronics in Agriculture. 2026;252:112075. doi:10.1016/j.compag.2026.112075
BibTeX
@article{Liu2026,
author = {Liu, Junbin and Xiao, Deqin and Tan, Zujie and Huang, Yigui},
title = {PA-SWE: a posture-aware sow weight estimation framework via skeleton-calibrated multimodal fusion},
journal = {Computers and Electronics in Agriculture},
year = {2026},
volume = {252},
pages = {112075},
doi = {10.1016/j.compag.2026.112075}
}
EUVIMED