Abstract:
Smart agriculture technologies are effective instruments for increasing farm sustainability
and production. They generate many spatial, temporal, and time-series data streams that, when
analysed, can reveal several issues on farm productivity and efficiency. In this context, the detection
of anomalies can help in the identification of observations that deviate from the norm. This paper
proposes an adaptation of an ensemble anomaly detector called enhanced locally selective combination
in parallel outlier ensembles (ELSCP). On this basis, we define an unsupervised data-driven
methodology for smart-farming temporal data that is applied in two case studies. The first considers
harvest data including combine-harvester Global Positioning System (GPS) traces. The second is
dedicated to crop data where we study the link between crop state (damaged or not) and detected
anomalies. Our experiments show that our methodology achieved interesting performance with Area
Under the Curve of Precision-Recall (AUCPR) score of 0.972 in the combine-harvester dataset, which
is 58.7% better than that of the second-best approach. In the crop dataset, our analysis showed that
30% of the detected anomalies could be directly linked to crop damage. Therefore, anomaly detection
could be integrated in the decision process of farm operators to improve harvesting efficiency and
crop health
