Automatic detection of reproductive behavior in male Japanese quail (Coturnix japonica) using accelerometers and neural networks.

R. MARÍN; L. BARBERIS; C. SIMIÁN; J. KEMBRO

Congreso; PSA Meeting; 2020

Tri-axial accelerometers placed on an animal measure the3-dimensional acceleration vector associated with bodymovements over time. When combined with machinelearning and data processing techniques, such as neuralnetworks, this methodology has the potential for classifyingthe recorded acceleration data into behavioral categories.Herein, we propose a system that implements the use of anaccelerometer attached to male Japanese quail as a usefulway for automatic detection of male reproductive behavior.Two different methods for attaching the accelerometer to the birds were also tested. Fifteen males and thirty femaleswere divided into one of three experimental groups: 1)control without accelerometer attached, 2) using anaccelerometer attached to a backpack (i.e. harness fitted by2 elastic fabric bands around the wings' base) or 3) using anaccelerometer attached to a patch made of fabric glued tothe back of the bird. All males were handled similarly andremained individually housed during a one-week perioduntil testing. The test initiated when a male was introducedinto the homebox of two female belonging to the sameexperimental group, during a 1-hour period. One cameraabove and one on the side of the box were used to recordbehaviors. From video-recording, a high resolutionethogram was performed defining all observable malebehaviors at a 1/15s resolution during the first 10-min oftesting (9000 data time points per bird). The number andduration of detected behavioral events were estimated.Accelerometer data was collected during the total 60-min oftesting. General linearized models were used to assessdifferences between groups in the most frequently observedbehavioral events, namely immobility, vigilance, shakes,exploration, walking, running, grabs, and mounts. In thevast majority of the variables evaluated no differences wereobserved between groups (P>0.05), including number anddurations of mounts. In a second stage, the high-resolutionbehavioral time series registered from video-recordingswere used first to train and then to validate a neuralnetworks, to automatically detect within the accelerometerdata the male reproductive events. Noteworthy, all displaysof reproductive behavior during the 1-hour testing periodwere detected with this method. Thus, the proposed systemis a first step towards automating the detection ofreproductive behaviors relevant for studies where visualobservations of video-recording are either not possible orimpracticable. In particular, this methodology could beuseful to assess male reproductive patterns over time withindifferent social and environmental contexts.