Automatic_Collembolan_Analysis

This repository contains code, example files and results files for the automatic image recognition of Folsomia candida species. Using this code it is possible to recognize F. candida individuals placed on a black background. The routine is used to count individuals in a sample, differentiate adult from juvenile individuals, and to get a data table with different size parameters for each individual animal. It is possible to calibrate the scale of the images, which allows for obtaining measurements in different units of size.

This routine was created using PyImageJ, a Python wrapper for ImageJ2, and scikit-learn, a machine learning library for Python that provides tools for data analysis and modeling.

Steps to run the analysis:

1. Create conda environment and install needed dependencies:

Create new environment:

conda create -n collembolanRecognition python=3.9

Activate the environment:

conda activate collembolanRecognition

Install dependencies:

conda install -c conda-forge pyimagej

conda install -c conda-forge scyjava

conda install -c anaconda scikit-learn

conda install numpy pandas matplotlib

2. Put the sample image in PNG format inside the img folder: image should be named like this: sample1.png

Folder structure looks like this:

├── collembolaAnalysis.py
├── img
│   ├── sample1.png
│   └── scale_calibration_results.csv
├── __init__.py
├── main.py

Sample image looks like this:

Source: Mallard et al. 2013

3. Run the collembolaAnalysis.py file

python collembolaAnalysis.py

After the run, new files are created, and the folder structure now looks like this:

├── collembolaAnalysis.py
├── img
│   ├── analyzed
│   ├── sample1.png
│   └── scale_calibration_results.csv
├── __init__.py
├── main.py
└── results
    ├── collembolan_size_analysis.xlsx
    ├── size_distribution_boxplot.png
    └── size_distribution_histogram.png

Inside the img folder, the analyzed folder is created, where analyzed image contains drawn polygons around original animals with the associated id number:

As a result of the polygons (i.e. each individual animal) size parameters measurement (diameter, area, circularity, roundness) the data table is created. This data table is used in machine learning method K-means clustering to derive two clusters: one representing juvenile animals and the other representing adult animals. It is also possible to deactivate juveniles from adult animals differentiation if the image only contains adult animals or juvenile animals.

In the results folder, size parameters of animals are stored in an Excel file. The file includes the following parameters for each animal: animal ID (from the previous analyzed image), Feret's diameter and area of each animal, standardized Feret's diameter, circularity, and roundness parameters, k-mean labels (0 or 1) - 0 being juvenile animal and 1 being adult animal, Sample Image Index - the index of image if there are more than 1 analyzed image, Feret's Diameter [μm]. The Excel file sheet summary contains basic descriptive statistics of given data (mean, minimum and maximum body size with standard deviation). Also, boxplot and histogram charts are created to visualize given data.

This routine can be used not only for automatization of the counting and measuring of collembolans, but for the automatization of microplastics particles quantitative and qualitative analysis. For the analysis of microplastics, the same logic applies as described above. Here are sample images that can be loaded into this routine:

The image is part of our research on soil microplastics. The microplastics on this image are artificialy made by using electric knife sharpener to hone HDPE bottle. Size distribution of the particles is roughly from 4 μm to 1 mm.

The analysis derived the following image:

The analysis derived the following size distribution charts:

It is also possible to differentiate between microplastics fragments and fibers by activating K-means clustering.

This routine is the part of the ongoing research of soil microplastics conducted at the Subdepartment of Quantitative Ecology, Department of Biology, University of Josip Juraj Strossmayer in Osijek, Croatia.

For any further information please contact: [email protected].