Automatic analysis of high, medium, and low activities of broilers with heat stress operations via image processing and machine learning

Heat stress is a major welfare problem in the poultry industry, altering broilers' activity levels. Advancements in image processing and machine learning provide opportunities to automatically quantify and analyze broiler activity. This study aimed to evaluate the effects of moderate heat exposure on broiler behavioral activity via image processing and machine learning. 132 Cobb 500 broilers were raised in 2 nutritional treatment groups, each with 3 replicates. The control groups were fed a basal diet, while the variation groups were fed a diet with 0.05 % 25-hydroxyvitamin D₃. All birds were raised under standard environmental conditions for 27 days before exposure to cyclic heat of 29.56 ± 1.34 °C and humidity of 76.97 % ± 5.98 % from 8:00-18:00 and thermoneutral conditions of 26.67 ± 1.76 °C and 80.23 % ± 3.05 % from 18:00-8:00. Birds were continuously video recorded, and the bird activity index (BAI) was analyzed by subtracting consecutive frames and summing up pixel differences. The treatment effect was analyzed using two-way ANOVA with a P-value < 0.05. K-means clustering was used to determine BAI as high, medium, and low levels. The result showed a significantly higher (P < 0.01) activity index in the variation group in contrast to the control. Absolute values of high and medium BAI were significantly lower with cyclical heating operations than those without heating operations. The BAI was also higher at the onset and end of the heating operations and moderately correlated to flock age (|r| = 0.35-0.45). The high, medium, and low BAI performed differently with different nutritional treatments, temperature ranges, and relative humidity ranges. It is concluded that the BAI is a useful tool for predicting broiler heat stress, but the prediction effectiveness could be influenced by bird age, diets, temperature, humidity, and behavior metrics.