Lipid monitoring of Chlorella vulgaris using non-invasive near-infrared spectral imaging

Abstract

Microalgal lipids are molecules of biotechnological interest for their application in sustainable food and energy production. However, lipid production is challenged by the time-consuming and laborious monitoring of lipid content in microalgae. This study aimed to predict the lipid content of Chlorella vulgaris cultivations based on non-invasively collected near-infrared (NIR) range hyperspectral data. A gravimetric analysis of total lipids was used as reference data (between 2 and 22% per dry weight) to compare three different models to determining the lipid content. A one-dimensional convolutional neural network and partial least squares models performed at a similar level. Both models could predict the lipid content of Chlorella dry weight with an error of 4%pt (root mean squared error). The index-based linear regression model performed the weakest of the three models, with the error of the prediction being 6%pt. Nile Red staining was used to visualise lipids on a microscope and lipid class analysis to resolve the lipid classes that explained most of the increase in lipids in Chlorella. A SHAP algorithm (SHapley Additive exPlanations) was used to analyse the wavebands of NIR spectra that were important for predicting the total lipid content. The results show that spectral data, when combined with an adequate algorithm, could be used to monitor microalgae lipids non-invasively in a closed system, in a way that has not previously been demonstrated with an imaging system.