Phenomics

A Computer Vision–Based Approach for Facial Temperature Monitoring and Non-Invasive Rectal Temperature Prediction in Dairy Cows

---

Abstract

In dairy cows, temperatures captured by infrared thermography (IRT) at facial regions are associated with invasive thermoregulation indicators such as rectal temperature (RT). However, the manual approach for thermographic data collection and processing is very labor-intensive, making it not feasible for real-time monitoring and large-scale phenotyping. Therefore, in this study, we investigated the feasibility of using a computer vision system (CVS) to automate temperature collection from dairy cows at different facial regions. The study was conducted at the UGA research dairy farm, located in Athens, Georgia. The cows were measured during 4 non-consecutive days in late June and early July 2025, between 10:00 and 16:00 h. A total of 1,600 thermal images were captured from 83 Holstein and Jersey cows' faces using a FLIR E54 camera. A climate sensor (Kestrel 5400AG) was installed inside the cattle chute to collect microenvironmental conditions. Additionally, RT was manually measured once for each cow during thermal image collection. Thermal images were manually annotated in six regions of interest (ROIs), including left eye, right eye, forehead, left ear, right ear, and muzzle base. Labelled images were used to train a deep learning algorithm (YOLOv11) for automated ROI segmentation using a leave-one-day-out cross-validation. For each testing set, a Python pipeline was used to extract temperature features from segmented ROIs. Linear correlation coefficients were computed between thermal features extracted from manual annotations and automatically via deep learning to estimate their equivalence. A Random Forest model was trained to predict RT based on the combination of CVS-assisted thermal features, baseline data (days in milk, breed, and lactation number), and climate variables. For segmentation, YOLOv11 achieved average precision ranging between 80.25% (left ear) and 92.90% (left eye), and mAP50 ranging between 79.78% (left ear) and 95.33% (left eye), across days and facial regions. Strong correlations were observed for the mean temperature in the forehead (0.97), left eye (0.95), and right ear (0.92) regions when comparing CVS-assisted and manually collected facial temperature features, suggesting reliable segmentation performance of YOLOv11. Maximum temperatures in the left eye (0.73 and 0.74) and left ear (0.71 and 0.47) measured via CVS demonstrated a moderate correlation with RT in Holsteins and Jerseys, respectively, indicating that these features can serve as potential non-invasive indicators of heat stress. The inclusion of automated thermal features substantially improved the prediction ability for RT compared to models trained with baseline and environmental variables only, with R2 increasing from 0.41 to 0.50, and MAPE decreasing from 0.91% to 0.82%. Results suggest that integrating IRT with CVS is a valuable approach for automatic, non-invasive monitoring of facial temperature, enabling large-scale assessment of heat stress indicators that could be leveraged in selection programs for thermal resistance in dairy cows.

Keywords: 2026

How to Cite:

Abdulrahman, T., Alves, A., Batista de Medeiros, B., Magalhaes Da Costa, R., Saha, C. & Tao, S., (2026) “A Computer Vision–Based Approach for Facial Temperature Monitoring and Non-Invasive Rectal Temperature Prediction in Dairy Cows”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286731. doi: https://doi.org/10.31274/wcgalp.24106

Rights: 1