Assessing sheep behavior through low-power microcontrollers in smart agriculture scenarios

Abstract

Automatic animal monitoring can bring several advantages to the livestock sector. The emergence of low-cost and low-power miniaturized sensors, together with the ability of handling huge amounts of data, has led to a boost of new intelligent farming solutions. One example is the SheepIT solution that is being commercialized by iFarmtec. The main objectives of the solution are monitoring the sheep’s posture while grazing in vineyards, and conditioning their behaviour using appropriate stimuli, such that they only feed from the ground or from the lower branches of the vines. The quality of the monitoring procedure has a linear correlation with the animal condition capability of the solution, i.e., on the effectiveness of the applied stimuli. Thus, a Real-Time mechanism capable of identifying animal behaviour such as infraction, eating, walking or running movements and standing position is required. On a previous work we proposed a solution based on low-power microcontrollers enclosed in collars wearable by sheep. Machine Learning techniques have been rising as a useful tool for dealing with big amounts of data. From the wide range of techniques available, the use of Decision Trees is particularly relevant since it allows the retrieval of a set of conditions easily transformed in lightweight machine code.

The goal of this paper is to evaluate an enhanced animal monitoring mechanism and compare it to existing ones. In order to achieve this goal, a real deployment scenario was availed to gather relevant data from sheep’s collar. After this step, we evaluated the impact of several feature transformations and pre-processing techniques on the model learned from the system. Due to the natural behaviour of sheep, which spend most of the time grazing, several pre-processing techniques were tested to deal with the unbalanced dataset, particularly resorting on features related with stateful history. Albeit presenting promising results, with accuracy over 96%, these features resulted in unfeasible implementations. Hence, the best feasible model was achieved with 10 features obtained from the sensors’ measurements plus an additional temporal feature. The global accuracy attained was above 91%. Howbeit, further research shall assess a way of dealing with this kind of unbalanced datasets and take advantage of the insights given by the results achieved when using the state’s history.