Benchmark of outlier detection methods for spectral data

UV/Vis spectrophotometers have been used to monitor water quality since the early 2000s. Calibration of these devices requires sampling campaigns to elaborate relations between recorded spectra and measured concentrations. Recent sensor improvements allow recordings of a spectrum in as little as 15 seconds, making it possible to record several spectra for the same sample. Spectrum repetitions provide new opportunities to detect outliers – a task that is difficult in non-repetitive spectra recordings. A well-executed outlier detection can e.g. result in a more accurate calibration of the spectrophotometer or an improved construction of a regression model. In this work, two methods are presented and tested to detect outliers in repetitions of spectral data: one based on data depth theory (DDT) and one on principal component analysis (PCA). Results show that the two methods are generally consistent in identifying outliers, with only small differences between the methods.