Abordagens de aprendizado de máquina para previsão de indicadores na bovinocultura de corte: auxílio em tomada antecipada de decisões

Abstract

The beef cattle industry is great of the activities with the greatest participation in animal protein. It's common, that the conventional model adopted in this livestock sector is divided into creates, recreates, and next, termination or herd reform. From one phase to another, bovines can be discarded or directed to an appropriate handling for their final destination. This decision is usually based on the interaction between the bovine's maintenance cost and its performance for growth, reproductive capacity, and meat cut or composition as an indication of its carcass yield. The adoption of new prediction technologies in beef cattle farming can help rural producers obtain better results in productive and economic terms, in addition to increasing the transfer and dissemination of technology in the livestock sector, through the development and integration of machine learning (ML). Thus, the aim was to present a proposal for a methodology using ML to support decisions in beef cattle production. The first part of the thesis describes, through a bibliographic review, the perspective of the use of ML in livestock. In the second part, it was proposed that using ML techniques can solve the issue of predicting animal performance pallowrely, allowing premature animal disposal. Thus, six different ML methods were compared (Linear Discriminant Analysis, Naïve Bayes, Tree Decision Classification and Regression, Support Vector Machines, Random Forest, and Neural Network) to identify the most accurate method to predict at 210 days the animals prone to discard at 450 days based on phenotypic indicators. A binary target trait (0/1) was created to represent animal disposal premature (1), using 3 phenotypic composite indices (PCI) created from the combination of growth characteristics, Kleiber index, scrotal perimeter, and carcass ultrasound. Animals with PCI equal to or less than 20% were considered prone to disposal prematurely. The accuracies of all methods were above 70%. According to different metrics, Neural Network and Random Forest methods performed best in prediction. In this sense, ML algorithms present themselves as a promising tool to improve farmers' decision-making. ML analysis would improve culling or redirection methods of handling an animal identified prematurely based on its performance at 15 months. The third part addresses a predictive data analysis to classify groups of cattle according to phenotypic performance (growth, reproductive, Kleiber index (IK), ultrasound, and economic indicators) in animals 10% inferior, intermediate group, and animals 10% superiors applying tree classification and regression, and multivariate analysis. The root mean squared error of the forecast was satisfactory, with a maximum of 5.42 and a minimum of 0.06. The values estimated by the forecast also showed satisfactory normality for the indicators. The groups created via machine learning corroborated the expected performance when analyzing their real data. Based on the minimum spanning tree, up to 450 days, the intermediate group was closer to the lower group by 10%, indicating that it is possible at that age to more accurately identify the higher animals. When ultrasonography is included (termination phase), the intermediate group is closer to the upper 10% group, that is, more facial to discriminate the lower group. Therefore, this methodology is promising to identify and differentiate in advance the performance of the animals, which can allow the cattle rancher to plan his management strategies.