Dung Kim Nguyen

Monitoring and estimating the density of microalgae in a closed cultivation system is a critical task in culturing algae since it allows growers to optimally control both nutrients and cultivating conditions. Among the estimation techniques proposed so far, image-based methods, which are less invasive, nondestructive, and more biosecure, are practically preferred. Nevertheless, the premise behind most of those approaches is simply averaging the pixel values of images as inputs of a regression model to predict density values, which may not provide rich information of the microalgae presenting in the images. In this work, we propose to exploit more advanced texture features extracted from captured images, including confidence intervals of means of pixel values, powers of spatial frequencies presenting in images, and entropies accounting for pixel distribution. These diverse features can provide more information of microalgae, which can lead to more accurate estimation results. More importantly, we propose to use the texture features as inputs of a data-driven model based on L1 regularization, called least absolute shrinkage and selection operator (LASSO), where their coefficients are optimized in a manner that prioritizes more informative features. The LASSO model was then employed to efficiently estimate the density of microalgae presenting in a new image. The proposed approach was validated in real-world experiments monitoring the Chlorella vulgaris microalgae strain, where the obtained results demonstrate its outperformance compared with other methods. More specifically, the average error in the estimation obtained by the proposed approach is 1.54, whereas those obtained by the Gaussian process and gray-scale-based methods are 2.16 and 3.68, respectively

This paper presents a low-cost system for efficiently monitoring density of microalgae in a closed cultivation system such as a photobioreactor. In fact, microalgal density can be accurately determined by manually counting methods such as the direct microscopic count technique. However, the manual approaches are cumbersome, time-consuming and impractical to be implemented in a closed cultivation system. Therefore, in the proposed monitoring system, microalgae are first proposed to be pumped from a culturing tank in to a sample container placed inside a dark box. A low-cost camera is utilized to capture images of microalgae through the transparent sample container under artificial light. It is then proposed to represent microalgal density through two average pixel values of red and green color channels of the corresponding image. Moreover, Gaussian process is exploited to statistically learn a data-driven model of microalgae density given the measured images. The learned model can then be used to effectively predict density of microalgae where only their corresponding image data is required. The proposed approach was evaluated in a real-world closed bioreactor system of culturing Chlorella vulgaris microalgae, where the model was trained by 100 images selected randomly from 125 ones. In 10000 random runs, accuracy of the estimated density results is about 8.6% (± 1.8%).