Automated Identification and Characterization of Particles in Holographic Images from Lake Geneva

This repository contains code and notebooks for my semester project, conducted at the EPFL Center for Imaging and LIMNC. The project focuses on detecting, segmenting, and classifying particles (e.g., plankton and sediment) imaged using the LISST-Holo2 underwater holographic camera.

The goal is to extract both handcrafted and learned features from particles, assess their structure using clustering, and classify them into Living vs. Non-Living categories using supervised learning.

---

Project Overview

Segmentation comparison

Our morphological segmentation pipeline preserves contours and texture without GPU inference.

Feature space visualization

UMAP projection of particle features — clusters highly overlap, indicating weak separability.

Classification performance

ROC curves for Random Forest, XGBoost, and LightGBM — limited separability between Living and Non-Living.

---

Repository Structure

File	Description
src/preprocessing_creating_ROI_NN.ipynb	Preprocessing and ROI creation for feature extraction (including DenseNet compatibility)
src/shape_analysis.ipynb	Extracts handcrafted shape, texture, and intensity features from segmented ROIs
src/classification.ipynb	Supervised classification using handcrafted features (Random Forest, XGBoost, LightGBM)
src/NN_features.ipynb	DenseNet121-based feature extraction and classification
src/dino_extraction_classification.ipynb	Self-supervised DINO feature extraction and classification
src/size_vs_no_size_filter.ipynb	Clustering analysis with and without area filtering
/report	Scientif report
/media	Illustrating figures

---

Requirements

You can install the required Python libraries using:

pip install -r requirements.txt

The main packages used include:

• numpy, pandas, scikit-learn, matplotlib, seaborn
• torch, torchvision
• umap-learn, scikit-image, albumentations

If you're using conda, we recommend creating a dedicated environment:

conda create -n lake_snow python=3.10
conda activate lake_snow
pip install -r requirements.txt

DINO Installation

This project uses DINOv2 (ViT-S/14) for self-supervised feature extraction.
Please follow the installation instructions from the official repository:
https://github.com/facebookresearch/dino

How to Run the Notebooks

The notebooks are intended to be run in the following general order:

1. Preprocessing

• preprocessing_creating_ROI_NN.ipynb: Prepares Regions of Interest (ROIs) for feature extraction.

2. Feature Extraction

• shape_analysis.ipynb: Extracts handcrafted features (shape, texture, intensity).
• NN_features.ipynb: Extracts deep features using DenseNet121.
• dino_extraction_classification.ipynb: Extracts self-supervised features using DINOv2 and runs the same classification pipeline as in classification.ipynb.

3. Analysis

• size_vs_no_size_filter.ipynb: Performs clustering on feature sets, with and without area filtering.
• classification.ipynb: Performs supervised classification (Living vs. Non-Living) using handcrafted features.

Results & Report

The full technical report is available in the report/ folder.

Data

Due to size constraints and EPFL policies, raw holographic .tiff files and annotated masks are not stored here. Please contact antoine.tissot-favre@epfl.ch for access if needed.

Acknowledgments

This work was supervised by:

• Prof. Edward Andò (EPFL Center for Imaging)
• Mallory Wittwer (EPFL Center for Imaging)
• Florian Aymanns (EPFL Center for Imaging)
• Natacha Pasche (LMINC)