Assessment of layer pullet drinking behaviors under selectable light colors using convolutional neural network

Light colors are important for poultry production performance, behavior, and well-being. However, drinking pereferences of layer pullets under light colors that can provide insights into welfare-oriented poultry managements remain unclear. The objectives of this study were to (1) develop a convolutional neural network (CNN) drinking behavior detector and (2) evaluate pullet drinking preferences for four light colors (white, red, green, and blue) in a lighting preference test system. The system consisted of four identical compartments with each containing a respective light color and two nipple drinkers, and pullets can move freely to make drinking choices between adjacent compartments. Three flocks of 20 Chinese domestic Jingfen layer pullets (54-82 days of age) were used for the test. The overall pullet drinking behaviors in each compartment were recorded by a camera atop each compartment. The recorded videos were converted to images for behavior analysis, and 5000 images of each drinker were used to train, validate, and test the faster region-based CNN drinking behavior detector. Daily time spent at drinkers (DTSD), percentage of simultaneously drinking pullets, and hourly time spent at drinkers under the four selectable light colors were analyzed based on the developed CNN detector. The results show that the detector had overall 88.2% precision, 88.7% recall, 89.4% specificity, and 89.1% accuracy on pullet drinking behavior detection using the testing set. The DTSDs (mean +/- s.e., minpullet(-1)d(-1)) were 13.2 +/- 1.1 under the white, 5.7 +/- 1.1 under the red, 3.5 +/- 1.1 under the green, and 17.0 +/- 1.1 under the blue. Less than two pullets choosing to drink simultanesouly at a drinker accounted for most of the time, and maximum number of simultaneously drinking birds was four. The pullets preferred to drink under the blue light the first six hours after the lights came on and under the white light within the last six hours before the lights went off. Overall, most pullets preferred to drink under the blue and white lights. It is concluded that the CNN-based behavior detector is a useful tool to detect pullet drinking behaviors, and the behavioral responses provide some insights into drinking management under different light colors to meet pullet drinking preference.