B1830
Title: Missing value imputation of sensor data for environmental monitoring
Authors: Thomas Decorte - University of Antwerp (Belgium) [presenting]
Steven Mortier - Antwerp University (Belgium)
Christian Suys - University of Antwerp (Belgium)
Tim Verdonck - KU Leuven and UAntwerpen - imec (Belgium)
Abstract: Over the last few years, sensor data has emerged as a crucial component in many operations generating massive spatio-temporal datasets. Nonetheless, sensor data frequently contains a (large) range of missing values, either due to systematic issues or inadvertent misoperations, which subsequently pose significant challenges during further analysis. Addressing missing observations in a collection of spatio-temporal sensor time series data involves accounting for both the temporal correlation between different timestamps of a single time series as well as the spatial correlation between the different time series or sensors. This aspect makes the missing value imputation of spatio-temporal data even more complex. Efficient methods are investigated for imputing the spatio-temporal sensor data of a large-scale environmental monitoring system consisting of over 4000 sensors for temperature and soil moisture monitoring. Methods based on spatial recovery as well as time series imputation and combinations of both are evaluated, with models ranging from the k-nearest neighbour (stKNN) and ARMA to the iterative imputing network (LSTM) and statistical methods (ST-MVL).
