Thermal signature: A method to extract characteristics from infrared thermography data applied to the development of animal heat stress classifier models

Author(s)

Rodrigues, Alex Vinicius da Silva ; Martello, Luciane Silva ; Pacheco, Veronica Madeira ; Sardinha, Edson Jose de Souza ; Pereira, Andre Levi Viana ; de Sousa, Rafael Vieira

• Document type: Journal article
• Source: Journal of Thermal Biology; v. 115, p. 8-pg., 2023-07-01.

Abstract

Recent studies have investigated the use of infrared thermography (IRT) to monitor body surface temperature and correlate it with factors related to animal welfare and performance. In this context, this work proposes a new method for extracting characteristics for the temperature matrix obtained using IRT data from regions of the body surface of cows which, if associated with environmental variables through a machine learning algorithm it generates computational classifiers for heat stress condition. IRT data were collected from different parts of the body of 18 lactating cows housed in a free-stall, monitored for 40 non-consecutive days, three times a day (5:00 a.m., 1:00 p.m., and 7:00 p.m.), during summer and winter, along with physiological data (rectal temperature and respiratory rate) and meteorological data for each time. The IRT data is used to create a descriptor vector based on frequency, accounting for temperatures for a pre-defined range, referred to in the study as 'Thermal Signature' (TS). The generated database was used for training and assessing computational models based on Artificial Neural Network (ANN) to classify heat stress conditions. The models were built using the following predictive attributes for each instance: TS, air temperature, black globe temperature and wet bulb temperature. The goal attribute used for supervised training was the heat stress level classification generated from the rectal temperature and respiratory rate values measured. The models based on different ANN architectures were compared through metrics of the confusion matrix between predicted and measured data, obtaining better results with 8 TS ranges. The best accuracy for classification into four heat stress levels (Comfort, Alert, Danger, and Emergency) was 83.29% using the TS of the ocular region. The classifier for two heat levels of stress (Comfort and Danger) obtained accuracy of 90.10% also using the 8 TS bands of the ocular region. (AU)

• FAPESP's process: 19/26828-1 - Prediction of thermal comfort level for dairy cattle: method based on machine learning modelling using climate chamber experiment data
• Grantee: Alex Vinicius da Silva Rodrigues
• Support Opportunities: Scholarships in Brazil - Master
• FAPESP's process: 18/05989-4 - Computational models based on Artificial Intelligence to assess the thermal stress of dairy cattle through non-invasive measurements
• Grantee: Alex Vinicius da Silva Rodrigues
• Support Opportunities: Scholarships in Brazil - Scientific Initiation