Training dataset 

The images were collected in the framework of the Belgian Lifewatch Research Infrastructure. During multidisciplinary campaigns, a number of fixed stations in the Belgian Part of the North Sea (BPNS) are visited on a monthly (onshore stations) or seasonal (offshore stations) basis. Samples are taken using a 55µm mesh size Apstein net and fixed in Lugol's iodine solution. In the lab, the samples are processed using a VS-4 FlowCAM model at 4X magnification targeting a particle size range of 55-300µm. The identification of the image data is done with the use of a CNN and followed by a manual validation step. Since May 2017, this dataset has provided micro- and phytoplankton observations, mainly covering diatoms, dinoflagellates and cilliates, for the Belgian Part of the North Sea (BPNS).

This dataset comprises a trainings datasplit of 337,514 images distributed across 95 classes, with each class containing a minimum of 100 and a maximum of 10,000 images. The goal of this dataset is to be able to facilitate model training, here we have organized the data into a standard split, with 80% allocated for training, 10% for validation, and another 10% for testing purposes. This dataset structure ensures a balanced representation and supports scientific rigor in subsequent analyses.

Technical details 

Data preprocessing 

Raw FlowCam output data is fully processed using in-house datapipelines, the VisualSpreadsheet software is only used for data acquisition during the lab run of the sample. Raw images and binary images are never saved during the FlowCam run, we only work on the image collages saved at the end of the run. Single images are cut from these collages using each image coordinates width and height pulled from the .lst file using in-house python code. The background of the images is not removed. These images are then predicted and annotated in-house at VLIZ.

Data splitting 

The training dataset is 80% used for training, 10% for validation and 10% for prediction. 

Classes, labels and annotations

The dataset comprises 337,514 images distributed across 95 classes, with each class containing a minimum of 100 and a maximum of 10,000 images. Taxonomic coverage of the dataset comprises mainly of diatoms, dinoflagellates and cilliates, but to a lesser extent also zooplankton and other protists.

Parameters 

The images are read using cv2.imread and the values are used as parameters.

Metadata Parameter Descriptions:

• image_path: The relative path to the image file showing the plankton or particle, usually structured by taxon or project folder.

• sample_datetime: The exact date and time (in YYYY-MM-DD HH:MM:SS.sss format) when the sample was acquired using the imaging instrument.

• flowcam_version: The software or hardware version of the FlowCAM instrument used to capture the image and process the sample.

• station: The sampling location code where the image was taken. These station codes typically refer to predefined geographic or monitoring points in the field.

• accepted_label: The final validated taxonomic label (e.g., genus or species) assigned to the organism or particle in the image, often based on expert review.

• accepted_aphia_id: The unique identifier (AphiaID) corresponding to the accepted taxonomic label in the World Register of Marine Species (WoRMS) database, which ensures standardized taxonomic reference.

• original_reference_id: A unique identifier (often a UUID) assigned to the image or sample in the original classification system (e.g., EcoTaxa), useful for traceability and linking back to the source record.

Data sources 

Images are collected during the monthly monitoring of phytoplankton communities in the Belgian Part of the North Sea during the LifeWatch multidisciplinary campaigns by FlowCam VS-4 benchmodel (Fluid Imaging Technologies, Yarmouth, Maine, U.S.A.).

Data quality 

All images are predicted and subsequently manually validated to ensure the quality of the trainingset.

Image resolution 

The size range imaged is 55-300µm. Images are acquired using a Sony XCD SC90 digital gray-scale camera. Images are during training of CNN resized to 100px by 100px.

Spatial coverage 

The data comes from a number of fixed stations in the Belgian Part of the North Sea (BPNS). 

Nine stations onshore are visited monthly:

Eight additional offshore stations are visited seasonally:

Temporal coverage 

The monitoring was initiated in May 2017 and has been running continuously every month.

Contact information 

For technical questions about training dataset, you can contact wout.decrop@vliz.be.