Accurate non-invasive quantification of astaxanthin content using hyperspectral images and machine learning

Abstract

Commercial cultivation of Haematococcus pluvialis to produce natural astaxanthin has gained significant traction due to its application in feeds, cosmetics and nutraceuticals. However, monitoring of astaxanthin content in cultures remains challenging and requires invasive, time consuming and expensive approaches. Here, we employed reflectance hyperspectral imaging (HSI) of H. pluvialis suspensions within the visible spectrum, combined with a 1-dimensional convolutional neural network (CNN) to predict the astaxanthin content (μg mg⁻¹) quantified by high-performance liquid chromatography (HPLC). This approach had low average prediction error (5.9 %) for samples cultured under the same conditions and was only unreliable at very low astaxanthin contents (<0.6 μg mg⁻¹). In addition, our CNN model outperformed single or dual wavelength linear regression models even when the spectral data was obtained with a spectrophotometer coupled with an integrating sphere. Different cultivation conditions to the ones used for the training of the CNN model affected the reflectance spectra of H. pluvialis cultures, leading to higher prediction errors (32.9 %). Transfer learning, a partial re-training of the model, fine-tunned the CNN to the variability observed in the new samples leading to a lower average prediction error (8.2 %). Overall, this study proposes the use of HSI in combination with a CNN for precise non-invasive quantification of astaxanthin in cell suspensions. Optimal performance is achieved when the training dataset accurately reflects the inherent variability of the target samples to be quantified.