Artificial intelligence based hyperspectral biomass estimator for cyanobacteria cultivation

Hyperspectral imaging combined with machine learning offers an innovative approach for biomass monitoring at laboratory and industrial scale, but a proof-of-concept linking hyperspectral data to biomass prediction remains limited. This study fills that gap by creating a dataset with 205 biomass measurements and 450 hyperspectral images from three cyanobacteria-rich microbiomes. Data were acquired using a compact push-broom camera, followed by image preprocessing to extract spectral information for training three machine learning algorithms based on either spectral or image data. The Fully Connected Neural Network model achieved the highest accuracy, predicting biomass levels with a mean absolute error of 37 mg/L (below 4 %). Notably, a simplified multispectral model using only three wavelengths reached comparable accuracy, highlighting the potential of low-cost multispectral systems. This study demonstrates the feasibility of hyperspectral and multispectral imaging for biomass estimation in cyanobacterial cultures and supports the development of real-time, non-invasive monitoring tools for photosynthetic bioprocesses.