Quantification of microalgae concentration based on multi-channel spectral fusion and 1D-CNN

Abstract Microalgae have been widely commercially cultivated, and their cell concentration is crucial for determining key cultivation parameters such as light intensity, temperature, and nutrient concentration. Absorption and fluorescence spectra are effective methods for detecting microalgal concentration. However, absorption spectra are weak and prone to interference at low concentrations, while fluorescence spectra are affected by the inner filter effect at high concentrations. To overcome these limitations, this study proposes a microalgal concentration prediction method based on a one-dimensional convolutional neural network (1D-CNN) that fuses multi-band light-emitting diode-induced fluorescence spectra and visible absorption spectra. We develop three fusion strategies: concatenation, channel-tacking, and dual-branch, and design three different 1D-CNN models for multispectral fusion, followed by performance comparisons. Experiments are conducted using CNN and three nonlinear machine learning models on multiple spectral datasets. The results show that multi-band fluorescence spectra fusion and asymmetric least squares processed absorption spectra significantly improve prediction performance. Using the fused spectral dataset for prediction yields the best results, with CNN performing notably better than other prediction models. Further comparisons of fusion strategies reveal that the channel-stacking fusion method yield the best performance, indicating that multichannel spectral fusion can improve the prediction accuracy. The model achieved a coefficient of determination ( R 2 ) above 0.989 and a root mean square error below 0.1000 on two microalgal test sets. Fluorescence and absorption spectral fusion, combined with deep learning, offer a feasible and cost-effective strategy for accurate algal biomass monitoring.