Predição do comportamento ingestivo de bovinos em pastejo ao longo do rebaixamento do dossel a partir de dados de acelerômetros

Abstract

The use of new technology has contributed to the increase in the efficiency of cattle production. Still, little of this technology has been used in pasture production, the basis of livestock in Brazil. With this in mind, the objective of this work was to evaluate different predictive models, validation strategies, and dataset compositions for the prediction of ingestive behavior of grazing cattle based on data generated by an accelerometer-type sensor. The experiment was carried out in intercropped pasture of Urochloa brizantha cv Marandu and Arachis pintoi. To test the change in ingestive behavior during lowering, the area was managed in a rotational grazing system, with an entry height of 25 cm and three exit heights (20, 15, and 10 cm). Observations took place over nine months, on non-consecutive days, for 12 hours a day. The predictive models used were: generalized linear regression (GLM), random forest (RF), gradient boosting (GB), and artificial neural network (ANN). The validations used were: holdout, leave-animals-out (LAO), leave-days-out (LDO), 10 cm leave-height-out (LHO10) and 25 cm leave-height-out (LHO25). Two datasets were used. The first was the PRO dataset, with grazing, rumination, and idle observations; the second was the PNP dataset, with grazing and non-grazing observations and an external database. Finally, the parameters used to evaluate the predictive models were: accuracy, error rate (for the PNP dataset), sensitivity, specificity, positive predicted value, and negative predicted value. For the PRO dataset, the best predictive model was the ANN, mainly in predicting grazing behavior, with an accuracy of 60.5% (LAO), 65.3% (LDO), 71.8% (holdout), 60 .5% (LHO10), and 63.2% (LHO25). In the PNP dataset, the best predictive models were the ANN, with an accuracy of 63.6% (LAO), 65.8% (LDO), 73.0% (holdout), 60.8% (LHO10) and 64 .9% (LHO25), and the RF, with an accuracy of 62.5% (LAO), 64.4% (LDO), 73.3% (holdout), 59.9% (LHO10) and 61.7% (LHO25). In general, the models were more efficient in predicting only two behaviors than in predicting three behaviors, mainly due to the difficulty in predicting idle behavior, with sensitivity below 26% in almost all validation strategies used. Another important point to consider is that the adopted validation strategy can interfere with the results of the evaluation parameters, as observed with the holdout, which had greater accuracy than other validation strategies when inflating the predictive models. The external dataset exposed the models to a new situation, where the holdout strategy was not superior to the others, with an accuracy of 57.5% (LAO), 59.4% (LDO), and 59.4% (holdout), showing the need to expose predictive models to new situations, such as the entry of new animals into the paddock and to different pasture structures.