微藻工厂化养殖为天然碳水化合物、 蛋白质等生产提供了重要途径, 但较高的养殖成本始终是限制微藻大规模商业化发展的瓶颈之一。 由于微藻生长速度很快且其胞内代谢信息时刻都在发生变化, 开发快速无损的微藻生长代谢监测手段用以实时获取微藻生产过程中感兴趣指标变化信息, 可以据此及时调整培养条件, 保障微藻高效优质生产。 关于微藻生长代谢信息快速无损检测的研究多集中在微藻油脂及其特性、 色素等方面, 对于同作为微藻重要营养成分的碳水化合物、 蛋白质等却少有报道。 本研究以斜生四链藻（Tetradesmus obliquus）为研究对象, 利用可见/近红外高光谱成像（HSI）技术结合化学计量学方法, 提出基于HSI的微藻碳水化合物和蛋白质反演判别方法。 对比研究标准化（autosacling）、 标准正态化（SNV）等12种预处理方法对原始高光谱数据的处理效果; 采用竞争自适应重加权采样算法（CARS）、 区间随机蛙跳算法（iRF）和模拟退火算法（SA）进行特征波段选择; 结合多元线性回归（MLR）、 偏最小二乘（PLS）、 支持向量机回归（SVR）以及随机森林回归（RFR）对微藻生物量、 碳水化合物和蛋白质含量进行反演判别。 结果表明, 斜生四链藻生物量预测模型采用矢量归一化（VN）预处理方式结合CARS-MLR算法效果最优, 决定系数（R2p）为 0.967, 剩余预测偏差（RPD）为6.212; 碳水化合物的最优预测模型为原始光谱（raw）结合iRF-RFR算法, R2p和RPD分别为0.995和36.156; 蛋白质判别模型采用WT预处理结合SA-RFR算法构建的效果最佳, R2p和RPD分别为0.909和10.116。 基于优化模型和HSI技术对藻液中各组分的空间分布及丰度进行了可视化展示。 研究结果有望为微藻工厂化养殖过程中生长信息的快速无损获取提供理论参考和技术支撑。

The industrial culture of microalgae provides an important way to produce natural carbohydrates and proteins. The high cost of cultivation is one of the main “bottlenecks” that influences microalgae industrialization development. Due to the high growth rate, the biomass and nutrients of microalgae vary rapidly. Therefore, a method that can -monitor the growth of microalgae and the dynamic information in the culture of microalgae would be of great necessity for the timely optimization of environmental parameters, thereby ensuring the efficient and quality production of microalgae. However, studies on the rapid and non-destructive detection of microalgae growth and metabolism information were mostly focused on lipids and their characteristics, with the other important components neglected, such as carbohydrates and proteins.In this study, T. obliquus was used as the research objects. A fast and nondestructive approach to estimate the carbohydrates and proteins concentration of T. obliquus in situ in a living environment was proposed based upon visible/near-infrared (VIS/NIRS) HSI system. Twelve data preprocessing approaches e.g. autoscaling and standard normal variate transform (SNV) etc., 3 feature selection methods including competitive adaptive reweighted sampling algorithm (CARS), interval random frog algorithm (iRF) and simulated annealing algorithm (SA), and 4 calibration models including multiple linear regression (MLR), partial least squares (PLS), support vector machine regression (SVR) and random forest regression (RFR) were applied to establish and optimize the estimation models. The results showed that vector normalization (VN) pretreatment combined with the CARS-MLR algorithm got the best performance on the biomass prediction of T. obliquus, with an R2p of 0.967 and RPD of 6.212. Raw spectra followed by the IRF-RFR algorithm performed the best for the carbohydrate of T. obliquus (R2p=0.996, RPD=36.156). Wavelet transform (WT) with SA-RFR obtained the best results for protein detection (R2p=0.909, RPD=10.116). Moreover, the visualization maps of these components spatial distribution and abundance in the microalgal liquid were obtained based on the optimal models. The overall results show that VIS/NIRS HSI is expected to be applied for efficient and high-quality production in microalgae industries.