Semi-supervised anomaly detection from Chlorella vulgaris cultivations using hyperspectral imaging

More evolved anomaly detection methods are needed to ensure efficient quality control of microalgae cultivations. This study aimed to determine whether non-invasively collected hyperspectral data can be used to indicate anomalies in Chlorella vulgaris cultivations. Three models of varying computational complexities were tested: isolation forest (iForest), one-class support vector machine (OC SVM), and neural network autoencoder. The models were trained in a semi-supervised way using 280 non-anomalous Chlorella spectra. test data included artificially contaminated cultures with Microcystis aeruginosa (4 spectra), nitrogen-depleted cultures (24 spectra) and non-anomalous Chlorella cultivations (43 spectra). The OC SVM was the most sensitive in detecting anomalies (AUC = 0.87 [0.79, 0.95], F1 = 0.91 CI [0.85, 0.98]), although the 95 % confidence intervals (CI) overlapped with the metrics of the other models. The model detected artificial contamination when the amount of Microcystis was around 1 % (biomass/ biomass) in the cultivation and nitrogen depletion after 3 days of nitrogen-free cultivation. The advantage of the semi-supervised training was that the models were able to learn about the normal Chlorella cultivations used as training data, and thus to classify unknown anomalies that deviated from the learned features. This may prove useful for detecting wider range of anomalies, but further testing is required to assess whether the other potential contaminants affect the spectra imaged in such a way that they differ from the normal. Non-invasive hyperspectral imaging together with the semi-supervised models provides a rapid indication method that could potentially give microalgae producers up-to-date information on cultivation quality.