Machine learning (ML) is a subset of artificial intelligence (AI) and computer science that is concerned with the use of data and algorithms to improve performance or to make accurate predictions [134]. Data science is a rapidly expanding area that relies heavily on ML. In data mining projects, data inputs and statistical methods are used to train algorithms to make proper divisions, characterizations, and predictions, uncovering key insights within projects. Decisions made based on these insights can then have a direct impact on key growth indicators in fields such as agriculture and allied sectors. The key outcome of ML is generalizability; the algorithm’s ability to correctly anticipate new data based on previously acquired rules [135].

Recently, a concept named Precision Livestock Farming (PLF) has been implemented to improve the farming process. In traditional livestock farming, the producer decides a factor based on his experiences, whereas in PLF, such decisions are based on quantitative data, such as liters of milk per milking, respiratory rate during the afternoon, behavior during heat stress, methane production, and many more aspects. In addition, quantitative data can be obtained in real time [136]. To study real-time complex and huge data, PLF depends on systems such as ML, control systems, and information and communication technologies [137]. ML has been used in different fields due to its versatility and ability to derive a model from available data [138]; however, because ML is not always the best match, traditional linear models, such as logistic regression, are still used to make predictions in some cases [139]. In some cases, different ML alternatives can be compared, but it can be difficult to predict which strategy will yield the best results [140]. Many ML models exist and may be suitable to predict the variable of interest. In nutshell, a trial-and-error strategy can be utilized to determine the most appropriate procedure for each prediction [141]. As mentioned previously, to process and discover abnormalities in the data that impact the production of animals, such as the effects of heat stress on livestock production, ML plays a major role in extracting huge data from various sensors put on a large population of animals and further processes it. ML can offer a scalable solution, utilizing data from different sources such as hardware sensors, e.g., pressure sensors, thermistors, infrared thermal imaging sensors, facial recognition machine vision sensors, Radio Frequency Identification, accelerometers, microphones, etc. [142]; functional data such as climate variables, weight estimates, physiological parameters (respiration rate, pulse rate, rectal temperature, and skin temperature), animal behavior, feed intake, and water intake assisted with biochemical and endocrine parameters.

The idea of introducing sensors on the farm is to provide decision-making information to the farmer. Through the GPS, in pasture-based dairy cows, it is easy to distinguish between grazing, resting, and walking through temporal positioning. Different models such as JRip, J48, and random forest all classified resting with an accuracy of 0.85 or more, while all models failed to accurately categorize grazing behavior (accuracy: 0.16–0.72). Additionally, researchers used GPS locations to forecast cow behavior and effectively detected transition moments between walking, grazing, and resting [143]. Similar behavior analysis studies were carried out using accelerometers mounted on the cow’s neck and leg [144]. Further, a neural network model was used on data from radiofrequency identification (RFID) sensors along with automatic milking system data to track cows. Recently, by combining background reduction and inter-frame difference models, Guo et al. [145] built a machine vision model for calf behavior recognition, and the detection rate was over 90%. These behavioral models can be implemented in heat stress studies to predict the changes in behaviors when animals are exposed to heat stress.

Several climate variables play a major role in causing heat stress, such as air temperature, solar radiation, relative humidity, and wind speed. Ranking these variables according to their impact on animals can be achieved through ML. In a study carried out by Gorczyca [146], where four models (penalized linear regression, random forests, gradient boosted machines, and neural network models) were utilized, a random forest model was accurate in predicting physiological responses (core temperature, skin temperature, and respiration rate) during heat stress. Further, these authors stated that air temperature has the largest effect on physiological responses, followed by solar radiation and by the interaction of air temperature and relative humidity during heat stress; wind speed and relative humidity were negligible heat stressors to cows. To assess the physiological responses, Gorczyca and Gebremedhin [147] used non-linear models such as neural networks and random forest and claimed these models are top predicting models for respiration rate, skin temperature, and vaginal temperature (R2: 0.61, 0.85, and 0.472, respectively).

In another study by Kim and Hidaka [148], infrared thermography (IRT) was used along with RGB (red, green, blue) images to observe breathing patterns in cattle. The Mask R-CNN algorithm was used to determine breathing rates and breathing patterns and to detect the region of interest, in this case, the nose. Temperature fluctuations around the nose during respiration were monitored through IRT to determine the respiration rate per minute in animals. The author proposed that the R-CNN algorithm has an accuracy of 76% in detecting cattle nose and infrared thermography, and a deep learning algorithm can be used to determine breathing patterns whenever animals are exposed to different stressors. Using these algorithms, farmers may determine when to deploy heat stress mitigation techniques by computing thresholds for environmental factors. One of the subsets of ML, neural networks were applied to assess the level of thermal stress in feedlot cattle, considering both weather and animal factors, by Sousa et al. [149] in Nellore cattle. The results suggested that the neural model has a good predictive ability, which had an R2 of 0.72, while the normal regression model had an R2 of 0.57. The results suggested that the neural model has a good predictive ability, with an R2 of 0.72, while the regression model yielded an R2 of 0.57. The neural model predicted thermal stress and was well correlated with the measured rectal temperature (94.35%), and it was significantly better than the THI method’s performance. Another non-invasive method of recognizing the stress in animals is through livestock vocalization.

Milk yield decreases particularly during heat stress, especially in high-producing dairy cattle, which have higher metabolic turnovers [150]. Benni et al. [151] predicted the response of milch cattle to heat stress, where they proposed a generalized additive model with mixed effects that was suitable for describing each cow’s response to crucial THI circumstances on milk production. Furthermore, they asserted that this analysis can contribute to improving herd management during heat stress conditions and identify the cows most likely to be suffering during heat stress to implement some specific actions for these animals in terms of cooling treatments, the enhancement of the feeding strategies, and attention to their specific welfare conditions. Another study by Bovo et al. [152] revealed that the Random Forest Model can detect a decrease in milk yield in cows due to heat stress effects caused by extremely hot temperatures. Indeed, the model’s average relative error in the forecasts is approximately 18% when considering a single daily yield but drops to just 2% when considering the total milk output during the test days.