Abstract
A new extraction-free and non-destructive method to evaluate pigment content (chlorophyll a, chlorophyll b, and total carotenoids) in microalgae cells is introduced. It relies on the use of a spectrophotometer and an optical diffuser, together coined as the opal glass method. First, the practical aspects of the technique are consolidated (ease of setting up, usage of disposable cuvettes, speed of data acquisition, …). Then, a database is created to link opal glass spectra to microalgae pigment contents obtained by a methanol-extraction method. Partial Least Squares (PLS - machine learning) and Convolutional Neural Networks (CNN - deep learning) are investigated to retrieve cell pigment content from opal spectra. Data preprocessing (filtration, augmentation/derivative-based feature engineering, signal scaling) is investigated in a systematic manner for both algorithms. In addition, two CNN architectures are tested, and hyperparameters are optimized for each possible combination. Overall, PLS shows merit, but sometimes offers some misprediction, resulting in a 10 to 15 Mean Absolute Percent Error (MAPE). On the contrary, the CNN achieves a stable sub-10 MAPE, making it the tool of choice to link opal glass spectra and cell pigment content. Finally, a learning curve analysis is carried out to demonstrate that the number of samples required to train the algorithms remains manageable (60-90 for the PLS, 80-150 for the CNN). Overall, this technique reduces pigment quantification from several hours (at best) to less than two minutes and opens the way to online reliable microalgal pigment content monitoring.