Predictive model based on artificial neural network for assessing beef cattle thermal stress using weather and physiological variables

The performance of feedlot cattle is adversely affected by thermal stress but the approach to assess the status of animal stress can be laborious, invasive, and/or stressful. To overcome these constraints, the present study proposes a model based on an artificial neural network (neural model), for individual assessment of the level of thermal stress in feedlot finishing cattle considering both weather and animal factors. An experiment was performed using two different groups of Nellore cattle. Physiological and weather data were collected during both experiments including surface temperatures for four selected spots, using infrared thermography (IRT). The data were analyzed (in terms of Pearson's correlation) to determine the best correlation between the weather and physiological measurements and the IRT measurements for defining the best body location and physiological variable to support the neural model. The neural model had a feed-forward and multi-layered architecture, was trained by supervised learning, and accepted IRT, dry bulb temperature, and wet bulb temperature as inputs to estimate the rectal temperature (RT). A regression model was built for comparison, and the predicted and measured RTs were classified on levels of thermal stress for comparing with the classification based on the traditional temperature-humidity index (THI). The results suggested that the neural model has a good predictive ability, with an R-2 of 0.72, while the regression model yielded R-2 of 0.57. The thermal stress predicted by the neural model was strongly correlated with the measured RT (94.35%), and this performance was much better than that of the THI method. In addition, the neural model demonstrated good performance on previously unseen data (ability to generalize), and allowed the individual assessment of the animal thermal stress conditions during the same period of day.