Search

Direct monitoring of wild animals' behavior is challenging and data tampering. Instrument the animals with collars that embeds sensors, such as tri-axial accelerometer and GPS, allows obtaining sufficient information for remotely classifying the performed activities. In this work is presented an accurate and human intelligible framework, designed leveraging the authors' skills in machine-learning and data analysis. The system covers all the steps required to accurately map the raw signals to the activities carried out, grouped in pre- processing, features extraction and selection, and classification phases. A case of study consists of a dataset collected by the Crofoot Lab at the Mpala Centre, in Kenya, instrumenting free-ranging Olive Baboons. This dataset provides both sensors time-series paired with respective activity labels and unlabeled ones. Labeled data was used to tune the parameters of the framework phases and to train and test the employed boosted- trees classifiers, while unlabeled ones were used for further system validations. The average accuracy obtained on a single activity is 94.5%. At best of the authors' knowledge, this is the first work that aims to solve the problem of direct human monitoring with such high accuracy, outperforming the state of the art by a lift about 10%. The main contribution of the proposed systems consists of the attention paid to the ethologist's needs. Together with the predictions, the framework also returns a ranking for the features considered, based on their importance in the decision-making process of the classifier. Therefore, the extracted features are consistent with the logical human path that the ethologist follows in performing direct monitoring. The produced framework has also been designed consistently with the ethogram structure to be easily interpretable and to allow activities classification at different aggregation levels. Unlabelled Image • Machine-learning algorithm enables automatic remote wild animals' monitoring. • Ad hoc features selection increase framework interpretability. • A classifier consistent with the ethogram ease decision-making understanding. • Inconsistencies detection algorithm makes the predictions robust to sensor failures • Framework performances are assessed analyzing a real case study. [ABSTRACT FROM AUTHOR]