Investigating quantitative approach for microalgal biomass using deep convolutional neural networks and image recognition

• Published in: Bioresource technology 2024-07, Vol.403, p.130889-130889, Article 130889
• Main Authors: Peng, Yang, Yao, Shen, Li, Aoqiang, Xiong, FeiFei, Sun, Guangwen, Li, Zhouzhou, Zhou, Huaichun, Chen, Yang, Gong, Xun, Peng, Fanke, Liu, zhuolin, Zhang, Chuxuan, Zeng, Jianhui
• Format: Article
• Language: English
• Subjects: Algal species; Algorithms; Artificial intelligence; Biomass; Image Processing, Computer-Assisted - methods; Machine vision; Microalgae - metabolism; Neural Networks, Computer; Noncontact detection; Rhodophyta - metabolism; Spirulina - metabolism; Yield quantification
• ISSN: 0960-8524; 1873-2976
• DOI: 10.1016/j.biortech.2024.130889

[Display omitted] •EffNet and ResNet excel in microalgae biomass quantitative monitoring.•Rhodophyta RGB trends are more suited for quantitative forecasts than Spirulina.•ResNet demonstrates greater generalization and robustness than EffNet. The effective monitoring of microalgae cultivation is crucial for optimizing their energy utilization efficiency. In this paper, a quantitative analysis method, using microalgae images based on two convolutional neural networks, EfficientNet (EFF) and residual network (RES), is proposed. Suspension samples prepared from two types of dried microalgae powders, Rhodophyta (RH) and Spirulina (SP), were used to mimic real microalgae cultivation settings. The method’s prediction accuracy of the algae concentration ranges from 0.94 to 0.99. RH, with a distinctively pronounced red-green-blue value shift, achieves a higher prediction accuracy than SP. The prediction results of the two algorithms were significantly superior to those of a linear regression. Additionally, RES outperforms EFF in terms of its generalization ability and robustness, which is attributable to its distinct residual block architecture. The RES provides a viable approach for the image-based quantitative analysis.