Automated classification of sample images collected by CEFAS Plankton Imager. Three classes: copepod, non-copepod and detritus.
Plankton plays an essential role in the global carbon cycle and carbon sequestration, regulating the exchange of carbon dioxide between the atmosphere, surface ocean and ultimately the seabed. Plankton is also used in global monitoring efforts providing reliable and sensitive indicators to climate change and ecosystem health.
As part of a Data Study Group (DSG) challenge organised between the Alan Turing Institute, the Centre for Environment, Fisheries and Aquaculture Science (CEFAS) and Plankton Analytics Ltd (see here), the participants contributed in the use of pretrained Convolutional Neural Networks (CNNs) with a ResNet-50 architecture to improve the accuracy of plankton classification at finer taxonomic levels compared to a baseline Random Forest (see the full report here).
In this notebook, we demonstrate how scivision facilitates the discovery of one of the trained ResNet-50 CEFAS DSG models for classifying plankton images into three classes: copepod, non-copepod and detritus. We pair the model with one of the matched data sources from the scivision data catalog, in this case a relatively small sample of images (n=26) extracted from the full test set (N=5863) used during the DSG challenge.
The notebook is designed to be launched from Binder.
- Click the Launch Binder button at the top level of the repository
You may also download the notebook from GitHub to run it locally:
- Open your terminal
- Check your conda install with
conda --version. If you don't have conda, install it by following these instructions (see here) - Clone the repository into your current folder:
git clone https://github.com/scivision-gallery/plankton-classification.git - Move into the cloned repository,
cd plankton-classification - Install the dependencies, in a new environment:
conda env create -f environment.yml - Activate the installed environment,
conda activate plankton-classification-scivision - Launch the jupyter interface of your preference, notebook,
jupyter notebookor labjupyter lab
A full conda environment can be found in environment_full.yml. If you are having issues creating the conda environment, this file contains details of all the packages and their versions which are required to run these notebooks, including secondary and tertiary dependencies.
This notebook was supported by the outcomes of the CEFAS DSG challenge in November 2021. The scivision team thanks the individuals and institutions involved in the Plankton DSG, in particular CEFAS for providing one of the trained models and sample images used in this notebook.
Data Study Group Final Report: Centre for Environment, Fisheries and Aquaculture Science (Cefas).