Marine microalgae are widely distributed in the ocean (Lu et al., 2021a) as part of the “blue carbon sink” that uses solar energy and dissolves CO₂ to produce oxygen as well as carbohydrates through photosynthesis (Lu et al., 2021b). They are thus involved in the global ocean–atmosphere carbon cycle to mitigate anthropogenic CO₂ emission, which is the leading cause of escalating climate change. Healthy and viable algal growth is critical to the prosperity of diverse marine ecosystems in the ocean and carbon capture, utilization, and storage. Also, species and quantities of microalgae are widely used as indicators for marine ecological environment monitoring and water quality evaluation worldwide (Gilbert et al., 1992; Domenighini and Giordano, 2009; Gordon and Leggat, 2010; Parmar et al., 2016; Naughton et al., 2020; Peter et al., 2021).

At present, the method of identifying and counting algal species is mainly through manual observation under a microscope, which is time-consuming and laborious and relies on their expertise, and algal samples are very likely to die in the process of collection and transportation, as shown in Figure 1A . The methods to monitor the health of algal cells at the single-cell level include chlorophyll determination, flow cytometry, and molecular biology. However, these methods rely on complex sensors and large instruments, which are not easily popularized in the general laboratory. It is challenging to conduct real-time detection when sampling in the field, as shown in Figure 1B .

In order to monitor and determine the health status of marine microalgae in situ, hardware and software integration is needed to enable real-time algal image acquisition, data analysis, and implementation of machine learning algorithms to recognize and classify the cells of marine microalgae. Microfluidic technology is a promising approach to recognizing microalgae and monitoring the health of microalgae at the single-cell level. It can achieve high throughput, has good biocompatibility, provides the ability to integrate with other methods, and requires a small sample size. Combining microfluidic technology with microscopic imaging can avoid the potential damage of labeled cells and enable dynamic cell detection, as shown in Figure 2 . The method would produce images of algal cells or videos of algal flowing, just enough to provide data for computer vision methods.

Compared to current manual microscopic identification, which has disadvantages such as high professional level requirements, discontinuity of classifiers, and being time-consuming, automatic marine microalgae identification by using computer vision methods (Coltelli et al., 2014; Promdaen et al., 2014; Giraldo-Zuluaga et al., 2018; Reimann et al., 2020; Barsanti et al., 2021; Cao et al., 2021; Park et al., 2021; Xu et al., 2021) can meet the needs of rapid monitoring and provide convenience for researchers in marine and environmental science. In this analysis of microalgal images, automatic localization and identification are expected to be achieved simultaneously, facilitating the downstream cell analysis. As the joint tasks of classification and localization, object detection can provide the basis for algae identification based on image information combined with biomorphological features.

Object detection is the task of precisely estimating the concepts and locations of objects in each image (Felzenszwalb et al., 2010; Zhao et al., 2019). Traditional object detectors (Viola and Jones, 2001a; Viola and Jones, 2001b; Dalal and Triggs, 2005; Felzenszwalb et al., 2008) based on sophisticated handcrafted features used to be mainstream methods before the popularity of deep learning. Thanks to the capability of learning robust and high-level feature representations of images, convolutional neural networks (CNNs) have been widely used in object detection. Two-stage detectors (Girshick, 2015; Girshick et al., 2015; He et al., 2015; Ren et al., 2017; Cai and Vasconcelos, 2018; Sun et al., 2020; Zhang et al., 2020a) achieved object region selection and detection in two steps while one-stage detectors (Liu et al., 2016; Redmon et al., 2016; Lin et al., 2017; Redmon and Farhadi, 2017; Redmon and Farhadi, 2018; Bochkovskiy et al., 2020; Law and Deng, 2020; Feng et al., 2021; Ge et al., 2021; Jocher, 2021) gave the class probabilities and position coordinate values of objects directly. Recently, some transformer-based methods (Carion et al., 2020) were proposed, showing their strong ability in object detection. Commonly used object datasets include Microsoft COCO (Lin et al., 2014) and Pascal VOC (Everingham et al., 2010).

Some existing related works on microalgal microscopy images applied such a deep learning-based method. Cao et al. (2021) used MobileNet (Howard et al., 2017) and SPP (He et al., 2015) to improve YOLOv3 while Park et al. (2021) trained the YOLOv3 model with Darknet53 backbone (Redmon and Farhadi, 2018). Cao acquired microalgal images using a Nikon Ts2-FL/TS2 fluorescence microscope and then used the data augmentation methods to obtain many images (Cao et al., 2021). The microalgal species in this dataset include Prorocentrum lima and Karenia mikimotoi. Park et al. (Redmon and Farhadi, 2018) collected 1,114 algal images with 3,663 objects collected by a microscope (Eclipse Ni, NIKON, Japan). These images were split into 10, 20, or 30 genera to compare the model’s performance with different numbers of objects for classification.

In this study, we have taken samples of six genera of microalgae commonly found in the ocean (Pinnularia, Chlorella, Platymonas, Dunaliella salina, Isochrysis, and Symbiodinium) and created an image dataset for them. We also collected the images of Symbiodinium in different physiological states known as normal, bleaching, and translating, and an image set mixing all microalgal genera described above. We provided many details about this dataset, such as producing process, the original resolution, and the total number of images and objects. Since we will subsequently apply this dataset to the competition, the number of images containing Pinnularia was intentionally set lower than the others to examine how well the participants handled the sample imbalance problem in the dataset. In addition, we added an image set of mixed samples to the test set, which was designed better to fit the algal samples in the field environment.

We hope to discover minimal differences between cells from a computer vision perspective and understand life’s heterogeneity, randomness, and synergy at the single-cell level. We gave a brief description of the dataset information and the annotation process. Additionally, we trained some classical or state-of-the-art object detection methods as baselines on the training set and got a test result on the test set. Researchers are welcome to develop their own marine microalgal object detection algorithms on this dataset and compare their results with these baselines. We sincerely hope that this dataset of microscopy images for marine microalga detection will boost related research on marine biology and help establish future real-time monitoring and water quality evaluation system of the marine environment.

This section specifically describes the workflow of building this dataset. The workflow of the dataset production includes data (water samples containing different genera of microalgae) collection, microscopic imaging, image annotation, annotation proofreading, and splitting of the dataset into the training set and the test set, as shown in Figure 3 . In addition, we also described the training environment and relevant evaluation metrics of the object detection models.

Object detection is a computer vision task that involves locating and identifying objects in an image or video. Two main object detection methods exist: single- and two-stage object detection algorithms. One-stage object detection methods perform object classification and bounding-box regression directly without using pre-generated region proposals. Two-stage object detection methods first generate region proposals using a separate network or algorithm, such as selective search or region proposal network (RPN), and then classify and refine each proposal using another network. In this study, we trained one-stage object detection algorithms including YOLOv5, YOLOv8, and TOOD, and two-stage object detection algorithms including Faster-RCNN, Cascade-RCNN, and Dynamic-RCNN.

The detection results based on mmdetection on the test set are shown in Table 1 . The evaluation metrics of the one-stage object detection algorithms are higher than the two-stage object detection algorithms. The indices of mAP (IoU = 0.5) of the Faster-RCNN series are only approximately 0.692 to 0.756. The Cascade-RCNN series, which uses cascade regression as a resampling mechanism to increase the IoU value of the proposal stage by stage so that the proposals resampled by the previous stage can adapt to the next stage with a higher threshold, have a higher detection accuracy. However, Cascade-RCNN with a larger backbone has a lower detection accuracy than the one with a smaller backbone. This may be because the larger model causes overfitting more easily. Dynamic-RCNN, which continuously adaptively increases the positive sample threshold and adaptively modifies the SmoothL1 Loss parameter, also achieves better results than Faster-RCNN. TOOD, a one-stage detection method that uses Task-aligned head and Task Alignment Learning to solve the problem of classification and positioning misalignment, achieves better results than two-stage methods.

We also trained YOLOv5 and YOLOv8 models, two of the most advanced one-stage object detection algorithms, and explored the detection performance of models with different model sizes. The detection results of YOLOv5 and YOLOv8 on the test set are shown in Table 2 , and the visualizations of detection results on the test set are shown in Figure 7 . The detection results of YOLOv8m for different classes is shown in Table 3 and Figure 8 . The mean mAP (IoU = 0.5:0.95) of YOLO series is higher than the other algorithms, which indicates that its prediction boxes can locate and match objects more precisely at progressively higher IoU thresholds. YOLOv8 replaces the C3 structure of YOLOv5 with the C2f structure and adjusts the number of channels for different scale models. It separates the classification and detection heads, and changes from Anchor-Based to Anchor-Free, using the TaskAlignedAssigner positive sample distribution strategy and introducing Distribution Focal Loss. Compared to YOLOv5, YOLOv8 models lead to some improvement in our dataset. We can also find that larger models do not perform better than smaller models. In addition to the characteristics of the dataset image itself, this may also be related to the small amount of data.

In this paper, we apply algorithms in computer vision to the multiple object detection and physiological state assessment of marine microalgae. We introduced a new dataset for marine microalgae detection in microscopy images and the building process, including data collection, microscopic imaging, image annotation, proofreading, and splitting. This dataset contains microalgal images of six genera, namely, Pinnularia, Chlorella, Platymonas, D. salina, Isochrysis, and Symbiodinium. Moreover, we also collected the images of Symbiodinium in different physiological states, namely, normal, bleaching, and translating, which can be an indicator of the situation of coral and its water environment. The dataset contains 967 images, among which 537 are in the training set and 430 are in the test set.

We trained and tested a number of classical or state-of-the-art models in our dataset, including one-stage object detection algorithms and two-stage object detection algorithms. Among these algorithms, YOLOv5 and YOLOv8 were able to perform the task with high accuracy for the classification and localization of microalgal cells. The main challenges of this task can be concluded as follows:

For future research, we will make a larger dataset as the next version and continue to develop stronger baselines for the microalgal detection tasks.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/supplementary material.

Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

SZ: Conceptualization, Methodology, Resources, Visualization, and Writing—Original Draft; JJ: Investigation, Software, and Writing—Original Draft; XH: Data Curation; PF: Project Administration and Funding Acquisition; HY: Writing—Review and Editing, Supervision, and Funding Acquisition. All authors contributed to the article and approved the submitted version.