Termografia infravermelha aliada ao aprendizado de máquina na predição de estro na suinocultura

Abstract

The objective of this thesis was to evaluate the use of infrared thermography as a tool for estrus prediction in swine production. In Chapter 2, a meta-analysis protocol was conducted to identify the main research methods used for estrus evaluation through temperature measurements at different anatomical regions, including the vulva, snout, and anus. Four scientific databases (Scopus, Web of Science, Science Direct, and PubMed) were consulted, applying inclusion criteria (full-text articles published in English within the last 15 years) and exclusion criteria (conference abstracts, reviews, theses, dissertations, and book chapters). At the end of the selection process, six studies were included, and the data were organized in spreadsheets using Microsoft Excel. Information regarding vulvar, anal, and snout temperatures, as well as animal characteristics such as breed, age, estrus cycle stage, and sampling time, was extracted. The results indicated that temperature measurements, especially infrared thermography of the vulvar region, are more sensitive to hormonal changes associated with estrus, whereas internal methods, such as rectal temperature measurement, present higher risks due to handling. However, when not correlated with behavioral analyses, the interpretation of temperature changes is limited by animal stress, rendering the assessment incomplete. Therefore, further studies evaluating different anatomical sites in association with predictive models are recommended to increase accuracy and reduce detection failures. In Chapter 3, in response to the findings of Chapter 2, changes in superficial body temperature, ocular region, ear tips, chest, back, vulva, and perianal area of sows during the estrus period were evaluated using an infrared thermographic camera in nine Large White × Landrace sows. Images were analyzed using FLIR Thermal Studio software. Mean comparisons were performed using ANOVA followed by Tukey’s test (p < 0.05). Supervised machine learning models, including random forest, conditional inference trees, partial least squares, and k-nearest neighbors, were tested, and model performance was evaluated using a confusion matrix. The orbital region showed significant differences between estrus states. According to the confusion matrix, the algorithm was able to predict estrus with 87% accuracy in the test set containing 40% of the data when agitated behavior was combined with orbital area temperature.