Explainable Artificial Intelligence and Computer Vision as Powerful Tools for Precision Livestock Farming

The PLFDoc doctoral school is dedicated to advancing precision livestock farming by integrating cutting-edge technologies into farm management. Within our research program, we work on several distinct projects that showcase the potential of artificial intelligence-driven solutions to enhance farm operations, with a focus on animal welfare while also considering the needs of farmers and consumers. Further, two distinct projects, where the University of Applied Sciences Upper Austria is involved, one with a focus on explainable artificial intelligence and another with a focus on computer vision are showcased.
On the one hand we work on an explainable artificial intelligence approach for predicting the onset of farrowing in sows, a critical event for improving animal welfare and optimizing farm management. This study leverages accelerometer data from ear tags to detect nest- building behaviour and predict the time to farrowing. By combining acceleration metrics with prepartum examinations and farm management data, we develop a two-stage model. The first stage detects the onset of nest-building, while the second stage predicts the remaining time until farrowing. Various methods, including cumulative sum, Bayesian estimation of abrupt change, seasonality, and trend, and NestDetect, our custom model, are compared for nest-building detection. Symbolic regression and deep learning techniques are used to predict the time to farrowing. For 82.6% of the sows, the start of nest-building behaviour was detected within a 48-hour window before farrowing. When nest-building was detected accurately, symbolic regression predicted the remaining time to farrowing with a mean absolute error of 9.4 hours, while deep learning models achieved a mean absolute error of 9.6 hours. Notably, the symbolic regression model provided interpretable results, underscoring the importance of transparency in artificial intelligence- based predictions. This work highlights how explainable models in precision livestock farming can offer comparable prediction performance to black-box models while ensuring transparency for data-driven interventions.
On the other hand, apart from interpretability, the application of artificial intelligence-based solutions in practice is often hindered by the costly data preparation. Currently, there is much interest in developing vision-based systems for the automatic monitoring of livestock. The algorithms these systems are based on often require large amounts of labelled training data. However, in practice it is often infeasible to prepare large amounts of data and hence only a small subset of the available data is labelled. To prevent performance degradations due to only limited availability of labelled training data, we investigate strategies for improving the performance of models trained on small datasets. Specifically, we present a method for increasing the efficiency of fine-tuning pose estimation models for pig monitoring. The proposed method allows pre-selecting the most useful video frames from a large dataset by estimating their utility in absence of any labels. Frames are deemed high utility if they target the existing model’s weaknesses. Thus, the labelling effort is constrained to the selection of high-utility frames, allowing the creation of high performing models with reduced data preparation effort. Experimental results show that the models fine-tuned on the selected high-utility frames outperform models fine- tuned on the same amount of randomly selected frames.
The presented projects provide insights into the diversity of the PLFDoc doctoral school and show the potential to enhance efficiency and animal welfare in pig husbandry. Apart from the shown projects, further work is done by three more PhD students at Vetmeduni Vienna (two students) and TU Wien (one student). These students also focus on the application of computer vision in various fields of Precision Livestock Farming and on explainable artificial intelligence.