Measurement artefacts and inhomogeneity detection
ETC/ACM Technical Paper 2011/8

Released: 2011/12/06: See the report

This paper is a review and evaluation of methods for statistical detection of two different types of inhomogeneities, (I) outliers and (II) structural changes or breaks, in air quality measurement time series in AirBase.
A literature study was carried out and five simple stochastic methods were selected and tested with air quality data: autoregressive lag-2 model, moving window (whole window) test, moving window (two-sided window) test, lag-1 differences and moving average filter.
A limited set of test time series with labelled inhomogeneities from AirBase for monthly and hourly data were prepared to evaluate the methods on their performance (number of correct detections) and robustness (sensitivity of parameters for different data sets). The performance was measured using the Jaccard’s coefficient and the robustness was estimated by applying the methods on a validation data set with the parameters estimated beforehand using the test data sets.

For outlier (type I) detection the moving window (whole window) test showed very good results on performance and robustness. For the structural changes (type II) it was not as clear as for the outliers. The most promising method was the moving average filter with a tendency to over-detection. However, the visual checks of the method’s results indicated that the detection method of the extremes in variance of the window averages (which is the indicator for a structural change) could be improved and might reduce the over-detection. Further tests are recommended to develop a more robust method which can be used for automatic inhomogeneity detection.

Reference to R-scripts, data and other materials used for the calculations:

Prepared by: Lydia Gerharz, Benedikt Gräler, Edzer Pebesma (IfGI, Uni. Münster, Germany) under subcontract of ETC/ACM Consortium institute RIVM.

Published by: ETC/ACM, December 2011, 54 pp.