Australian Phytoplankton Microscopy Machine Learning Dataset

Full description

The Australian Phytoplankton Microscopy Machine Learning Dataset is an annotated collection of laboratory microscopy imagery developed to support the training and evaluation of deep learning object detection models for microalgae identification. The collection comprises ten datasets derived from the CSIRO Australian Phytoplankton Microscopy Dataset (CAPMD) source imagery. The collection spans three optical contrast modes (bright field, differential interference contrast, phase contrast), four objective magnifications (200X, 400X, 630X, 1000X) and two microscope platforms (Zeiss Axio Observer, Zeiss Axio Plan). The principal dataset is a multi-species training set covering ten species, most of them harmful algal bloom (HAB) taxa of monitoring interest, imaged at 630X under bright field microscopy on the Zeiss Axio Observer. Nine additional single-species datasets document Heterosigma akashiwo (strain CS1197) under controlled variation of optical contrast mode, objective magnification, and microscope platform, enabling quantification of cross-condition model transferability. Five datasets are provided with train, validation, and test partitions for direct model training. The remaining five serve as held-out cross-condition evaluation targets. All datasets are shipped in YOLO detection format, accompanied by the source segmentation polygons in COCO format, with an additional MMDetection-format copy of the multi-species dataset provided to support reproduction of the Faster R-CNN and Deformable DETR results reported in the companion paper.
Lineage: The annotation source imagery was drawn from the CSIRO Australian Phytoplankton Microscopy Dataset (CAPMD) 2021, 2022 and 2023 collections, an archive of approximately 270,000 laboratory microscopy frames acquired at the CSIRO Battery Point site in Hobart, Tasmania. Live monoculture samples from the Australian National Algae Culture Collection (ANACC) were imaged on a Zeiss Axio Observer inverted microscope (ZAO) with a Zeiss AxioCam HRc Rev. 3 detector and, for a subset of samples, on a Zeiss Axio Plan upright microscope (ZAP) with a Zeiss Axiocam 506 color detector. ZAO samples were held in coverglass-bottom petriplates and ZAP samples on slides beneath coverslips, and each sample was assessed for quality before acquisition. Each sample contributed short videos of sequential frames acquired with a focal rolling protocol that swept the objective through the sample depth. Highly motile cells were immobilised with 1% Tylose, which raises medium viscosity without altering the optical properties of the sample. Species coverage emphasised HAB taxa of monitoring interest, complemented by non-HAB strains for taxonomic and pigment-system breadth.

Proprietary Zeiss CZI acquisition files were decoded to individual JPG frames at 90% quality with the Marimba open-source Python framework (Jackett et al., 2025). Annotation was undertaken by trained domain experts in the Computer Vision Annotation Tool (CVAT), with each microalgae cell delineated by a segmentation polygon, drawn manually in the early stages and with the Segment Anything Model (SAM) integration in later ones. Because all samples were monocultures, every annotated cell was unambiguously attributed to its sample's strain. Annotation effort was guided by a target of 1,000 annotations per strain, with periodic quality-control reviews across annotators and over the duration of the project.

Ten machine-learning-ready datasets were assembled. The multi-species dataset (ZAO_X630_BF) contains 6,386 images and 15,591 annotations across ten species. Nine single-species datasets target Heterosigma akashiwo (strain CS1197): eight pair the three contrast modes with the 200X, 400X and 630X objectives of the ZAO, and the ninth is a ZAP 1000X bright field variant for cross-instrument comparison. Five datasets were partitioned into training, validation and test sets at a 70/15/15 target ratio with a sequence-aware, class-balanced splitting algorithm that allocates whole video-derived frame sequences to the test partition, so no test sequence is seen in training. The remaining five are cross-condition evaluation targets held entirely in the test partition.

Training and evaluation are reproducible from the scripts distributed with the collection. YOLOv8 training is deterministic under a fixed seed and reproduces bit-identically across independent runs on different hardware. The two OpenMMLab architectures are not bit-reproducible, because compiled CUDA kernels in mmcv do not honour PyTorch's deterministic-algorithm setting and Deformable DETR uses an operation with no deterministic CUDA implementation, so their results are reported as means over ten independent replicates. On the multi-species dataset, scored with the COCO API for all three architectures, YOLOv8 achieves an overall mAP of 0.914, Faster R-CNN 0.864 (standard deviation 0.006) and Deformable DETR 0.826 (standard deviation 0.008).

All ten datasets are provided in YOLO detection format, pairing JPG images with per-image label files in train, val and test subdirectories (test only for the evaluation-only datasets) alongside a data.yaml descriptor. An MMDetection-format copy of the multi-species dataset pairs the same 6,386 images with COCO-format train, val and test annotations. The segmentation polygons from which the YOLO boxes were derived are provided separately as one COCO JSON file per split for each dataset, covering all 30,283 annotations and referencing the imagery by filename.