Deep Learning Framework for Sow Posture Classification from Depth Images: Comparison with Data Transformation Techniques, Models, and Cross-validation

Authors

• Md Towfiqur Rahman, University of Nebraska-LincolnFollow
• Tami M. Brown-Brandl, United States Department of Agriculture, Agricultural Research Service, US Meat Animal Research CenterFollow
• Gary A. Rohrer, United States Department of Agriculture, Agricultural Research Service, US Meat Animal Research CenterFollow
• Sudhendu R. Sharma, University of Nebraska-Lincoln
• Suzanne M. Leonard, North Carolina State University
• Yeyin Shi, University of Nebraska-LincolnFollow

ORCID IDs

Rahman https://orcid.org/0000-0001-9520-2753

Brown-Brandl https://orcid.org/0000-0002-0874-8035

Sharma https://orcid.org/0000-0003-1996-673X

Shi https://orcid.org/0000-0003-3964-2855

Document Type

Article

Date of this Version

2026

Citation

Smart Agricultural Technology (2026) 15: 102388

doi: 10.1016/j.atech.2026.102388

Comments

Open access

License: CC BY-NC-ND 4.0

Abstract

Piglet preweaning mortality (PWM) in the United States averages ≈14–15%, with sow overlaying causing about one third proportion of these losses. This research aimed to develop and evaluate deep learning models to classify six sow postures using depth images to monitor behaviors linked to overlaying risk. Top-down depth images were captured with Kinect V2^® cameras at 10 frames min^-1 for five consecutive days (2 days before to 2 days after farrowing), yielding 26,506 training images from 18 sows, 17,901 testing images from 12 sows, and 4,697 additional images from three sows in diagonal stalls for external validation. Three transfer learning architectures (YOLOv11m-cls, ResNet-50, and Inception v3) were trained and evaluated on two different depth data trans- formed images (grayscale and Jet-colormap). Jet colormap images consistently outperformed grayscale, with YOLOv11m-cls achieving the highest accuracy (precision = 0.98, recall = 0.98, F1 = 0.98). External validation confirmed robust performance in diagonal stalls (F1 = 0.94), and cross-validation across stall designs and heat- lamp configurations demonstrated strong generalizability of the models. These findings show that depth imaging combined with deep learning provides a reliable, non-invasive method for sow posture classification. Such model can generate continuous behavioral data streams to improve understanding of postural transitions, stall design, and reduce piglet mortality.