Ocular infrared thermography combined with stacked ensemble machine learning enables contact-free rectal temperature prediction in Murrah buffaloes

Introduction Accurate core body temperature monitoring is central to health assessment in buffaloes, yet conventional rectal thermometry demands physical restraint and direct animal contact, causing stress to animals and posing injury risk to handlers. Methods This study developed and validated a contact-free prediction framework combining ocular infrared thermography with stacked ensemble machine learning, using medial canthus temperatures of both eyes alongside environmental variables as predictors of rectal temperature in 471 adult female Murrah buffaloes. Normalisation parameters were derived solely from the training set. Eight supervised algorithms were trained under five-fold cross-validation; the five best-performing learners were integrated into a stacked ensemble with a penalised linear meta-learner. A 48-h continuous monitoring experiment ( n = 10) further evaluated time-series concordance. Results On the independent test set, the stacked ensemble achieved the strongest performance (RMSE = 0.2089, 95% CI: 0.179–0.240; R 2 = 0.6152, 95% CI: 0.467–0.718), with random forest the best single learner. Clinically, 96.8% of predictions fell within ±0.5 °C of measured rectal temperature. Ablation analysis showed left eye temperature alone explained 36.6% of variance, rising to 56.0% with environmental variables and 61.5% with the full feature set. In continuous monitoring, LT, RT, and time were significant predictors (marginal R 2 = 0.651; conditional R 2 = 0.763), and Bland–Altman analysis confirmed a consistent mean bias of 1.5 °C between rectal and left eye readings. Discussion Ocular medial canthus thermography combined with ensemble machine learning offers a reliable, welfare-friendly alternative to rectal thermometry, with practical potential for integration into precision livestock health monitoring systems.