PlanktoShare Classifier

[paper] [dataset]

The PlanktoShare classifier predicts different plankton and non-plankton classes from data captured by the Plankton Imager (PI-10) sensor.

Getting Started

Data set-up

1. Download model weights from PLACEHOLDER. Two options are available, the ResNet50-detailed being more extensive with 49 different possible classifications, and the OSPAR model predicting 12 classes. Store these into /models/
2. Store your raw, unaltered Pi10-data into a preferable location. We recommend storing it in /data/, but can be stored in any accessible location using the argument --source_dir
3. For map creation, download the "EEA coastline for analysis" from the European Environment Agency. Store into /data/
4. For map creation, download the "Marine and land zones: the union of world country boundaries and EEZ's (version 4)" from Marineregions.org. Store into /data/

Anaconda set-up

# install the repository
git clone git@github.com:geoJoost/planktoshare.git

# Setup the environment
conda create --name plankton_imager

conda activate plankton_imager

conda install pip

pip install fastai

# IMPORTANT: Modify this installation link to the correct CUDA/CPU version
# Check the CUDA version using `nvidia-smi` in the command-line
# If no CUDA is available, use the CPU installation; Be aware that this is significantly slower and discouraged for larger datasets
# See: https://pytorch.org/get-started/locally/
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

conda install -c conda-forge pandas numpy polars seaborn xlsxwriter chardet geopandas python-docx memory_profiler pyarrow fiona pyproj

Usage

# To start the entire pipeline, navigate to your working directory
cd PATH/TO/WORKING_DIRECTORY

# Run the classifier
# See options below
# Not implemented yet
python main.py --source_dir data/YOUR_DATA_PATH --model_name ResNet50-detailed --cruise_name SURVEY_NAME --batch_size 300

# For more detailed options, see `main.py`

Options available in main.py:

• source_dir: This should be the path to your data folder directly from the Pi-10. It is recommended to store this within the repository in /data/.
• model_name: This corresponds to the model to use for inference. Options available are: ospar to use the OSPAR classifier (12 classes), or ResNet50-detailed to use the ResNet50 model which predicts 49 different plankton and non-plankton classes.
• cruise_name: This is used for intermediate outputs and for generating the final report. Any string is accepted without any spaces in the name, use '-' or '_' instead.
• batch_size: Number of samples to use within inference.py. This is highly dependent on the available memory within your PC/HPC. Default value of 32 is recommended for local machines.

Dataset Requirements

Use the original dataset structure as provided by the PI-10 imager without modifications.

Raw

CRUISE_NAME
├── 2024-06-24
│   ├── 1454.tar
│   ├── 1458.tar
│   ├── 1459.tar
│   ├── 1500.tar
│   ├── 1510.tar
│   ├── 1520.tar
│   ├── 1530.tar
│   ├── 1540.tar
│   ├── 1550.tar
│   ├── 1600.tar
│   ├── 1610.tar
│   ├── 1620.tar
│   ├── 1630.tar
│   └── 1640.tar
├── 2024-06-25
│   ├── 0000.tar
│   ├── 0010.tar
│   ├── 0020.tar
│   ├── 0030.tar
│   ├── 0040.tar
│   ├── 0050.tar
│   ├── 0100.tar
│   ├── 0110.tar

Untarred

CRUISE_NAME_UNTARRED
├── 2024-06-24
│   ├── untarred_1454
│   │   ├── Background.tif
│   │   ├── Bubbles.txt
│   │   ├── Cameralog.txt
│   │   ├── HitsMisses.txt
│   │   ├── RawImages\pia7.2024-06-24.1454+N00000000.tif
│   │   ├── RawImages\pia7.2024-06-24.1454+N00000001.tif
│   │   ├── RawImages\pia7.2024-06-24.1454+N00000002.tif
│   │   ├── RawImages\pia7.2024-06-24.1454+N00000003.tif
│   │   ├── RawImages\pia7.2024-06-24.1454+N00000004.tif

Inference Pipeline

The PlanktoShare repository automates the processing of PI-10 data using custom classifiers. The inference script performs the following steps:

1. Iterate through directory to process each .tar file.

2. Temporarily extracts all .tif images from each .tar.

3. Uses the classifier defined by the model_name argument to classify images.

4. Detects and discards corrupted images.

5. Generate outputs For each 10-minute bin, creates:

   • Detailed CSV: Includes per-image metadata:
     • Image details (filename, datetime, EXIF geodata)
     • Cruise information (cruise name, instrument code)
     • Model predictions (class ID/label, confidence scores)
   • Summarized CSV: Provides aggregated statistics:
     • Total predicted images per class
     • Density and summary statistics (e.g., average confidence)

6. Stratified random sampling (n=100) per class for manual validation and creating training data.

7. Automatically generates a summary report (examples in /reports/).

Future implementations

• Remove FastAI implementation

Known errors

• Error in learn.load(MODEL_FILENAME, weights_only=False) can be caused in older PyTorch versions. In this case, simply remove the weights_only argument.

---

If you use this code or dataset, please cite our paper. For questions, feedback, or collaborations, feel free to contact us.