Livestock behaviour forecasting via generative artificial intelligence
Regina Eckhardt, Reza Arablouei, Aaron Ingham, Kieren McCosker, Heinz Bernhardt
Abstract
Recent advancements in sensor technology and generative artificial intelligence (AI) are transforming precision livestock farming by enhancing behaviour monitoring and predictive analytics. This study examines the effectiveness of Transformer-type generative AI models in predicting cattle behaviour profiles and imputing missing data from collar accelerometer readings collected during two trials in Queensland, Australia, in 2022 and 2023, alongside climatic data. Each trial involved 60 cattle equipped with collars that classified six core behaviours: grazing, ruminating, walking, resting, drinking, and other over five-second time windows. Hourly behaviour profiles were constructed for each animal and experiment day by aggregating the behaviour predictions over every calendar hour, representing the time spent on each behaviour within each hour. Subsequently, four Transformer-type models (i.e., standard Transformer, Informer, Reformer, and Autoformer) were trained on the hourly behaviour profile data to predict behaviour profiles of the next 24 hours for each animal. Among the considered models, Autoformer showed the highest predictive accuracy when including climate data, achieving a mean absolute error (MAE) of less than 5.5 minutes, while the next best model had an MAE of approximately 6 minutes. For imputing missing data, the standard Transformer outperformed traditional imputation methods, with an MAE of less than 30 minutes over 24 hours, compared to 40 to 70 minutes for traditional methods (mean, median, and linear interpolation). These results highlight the potential of generative AI, particularly Autoformer and Transformer, to enhance predictive accuracy and data imputation in livestock management, thereby supporting regulatory guidance for data-driven decision-making and improved farming practices.