Adsa Logo White Adsa Title White

Applying machine learning on feeding behavior data for estrus detection in dairy heifers.

L. G. R. Pereira



Join L. G. R. Pereira on this page for a live text chat!
6:00 PM - 8:00 PM GMT


Applying machine learning on feeding behavior data for estrus detection in dairy heifers.
F. C. Cairo1, L. G. R. Pereira*2, M. M. Campos2, T. R. Tomich2, S. G. Coelho3, C. F. A. Lage3, A. P. Fonseca3, A. M. Borges3, J. R. R. Dorea4. 1Universidade Estadual do Sudoeste da Bahia Itapetinga, BA, Brazil, 2Brazilian Agricultural Research Corporation � Embrapa Juiz de Fora, MG, Brazil, 3Universidade Federal de Minas Gerais Belo Horizonte, MG, Brazil, 4University of Wisconsin-Madison Madison, WI.

The recent advances of sensor technology have allowed accurate estrus prediction using animal behavior information. The variables generated by Electronic feed and water bins have not been explored as predictive attributes for the development of models for estrus detection. The objectives of this study were (1) to quantify the effect of estrus on feed intake and animal behavior (feeding and drinking); and (2) to develop and evaluate predictive approaches to detect estrus using electronic feed and water bins data. Feed intake, animal behavior, and estrus events (n = 99) were measured in 57 Holstein � Gyr heifers (374 � 21.2 kg). Previous to each estrus event, the total feed intake (as-fed basis), number of feed bins visits, number of water bins visits, time spent eating and time spent drinking water were computed. Three predictive approaches were evaluated: logistic regression (LR); artificial neural network (ANN); and random forest (RF). Twelve covariate sets were established to (ii.a) evaluate the prediction quality for estrus detection when long (0 to −174 h) or short (0 to −24 h) time series were used as predictors; (ii.b) to evaluate the ability of models to predict estrus 6 and 12 h in advance; and (ii.c) to evaluate the predictive quality for estrus detection when only feeding and drinking behavior data (without intake variables) were included as predictors. All variables obtained by electronic bins change on estrus day compared with previous days. All predictive approaches analyzed with and without the feed intake variable were accurate for estrus detection. The short time series (24h) before estrus is satisfactory for estrus detection. The prediction of estrus in advance with 6 and 12 h reduced the accuracy and stability of the models. ANN models, RF and LR showed an accuracy of over 80%, indicating the possibility of predicting estrus at 06 h in advance. The exclusion of feed intake data does not reduce the accuracy, sensitivity, and specificity of models for estrus detection, indicating the possibility of developing new sensor-based devices that allow estrus detection.

Keywords: artificial neural network, heat detection, precision livestock.