Evaluating Modelling Techniques for Cattle Heat Stress Prediction

Researchers have traditionally predicted animal responses by means of statistical models. This study was conducted to evaluate modelling techniques. One hundred and twenty-eight feedlot heifers were observed during a 2-month period during the summer of 2002. Respiration rate and surface temperature were taken on a random sample of 40 animals twice a day. Five different models (two statistical models, two fuzzy inference systems, and one neural network) were developed using 70% of this data, and then tested using the remaining 30%. Results showed that the neural network described the most variation in test data (68%), followed by the data-dependent fuzzy model (Sugeno type) (66%), regression models (59 and 62%), while the data-free fuzzy model (Mamdami type) described only 27%. While the neural-network model may be a slightly better approach, the researcher may learn more about responses using a fuzzy inference system approach. For all models tested, respiration rate is over-predicted at low stress conditions and under-predicted at high stress conditions. This suggests that all models are lacking a key piece of input data, possibly the accumulative effects of prior weather conditions, to make an accurate prediction.