Non-invasive measure of heat stress in sheep using machine learning techniques and infrared thermography

Abstract Heat stress (HS) leads to altered sheep behavior, physiological, and biochemical processes which negatively affects their welfare and performance. While suitable strategies are needed to ameliorate the impacts of HS in sheep, it is equally important to accurately and non-invasively measure HS. Traditionally, rectal temperature (RT) is considered an indicator of thermal balance and is used to assess the impacts of hot conditions on sheep. However, measuring RT itself can be a stressor as it often requires restraining of the animals. The main objective of this study was to establish whether a combination of infrared thermography (IRT) and machine learning techniques can be applied to predict sheep RT when subjected to HS. Thermal images and RT were taken twice weekly from Dorper, and 2nd Cross (Poll Dorset X (Border Leicester X Merino)) lambs (n = 24/breed, 4-5 months old), for two weeks. Sheep were randomly allocated to either (i) thermoneutral (TN; 18–21 °C, 30–50% relative humidity (RH), n = 12/group) or (ii) cyclic HS treatments (28–40 °C, 40-60% RH, the cycle comprised of high temperatures 38−40 °C between 0800 and 1700 h daily and 28 °C, 30-40% RH maintained overnight). The head was selected as the region of interest because of less wool cover; specifically, the IRT of forehead, eye, ear, nostril, and face locations were measured. Artificial neural network (ANN) models were developed using three different backpropagation algorithms with temperature-humidity index (THI), and IRT temperatures as inputs and RT measured manually as targets. Results showed that the forehead and eye IRT temperatures had the highest correlation (P