IMG_SEGMENTATION – Satellite Image Segmentation Project

• Date: 2026-01-21
• Status: In Progress
• Author: Colin MANYRI
• License: MIT – Copyright (c) 2026 Colin MANYRI
• Version: 10.0.26

Project Goal

This project implements several complete pipelines for satellite image segmentation, including:

• Image preprocessing
• Feature extraction
• CNN-based model training (e.g., U-Net)
• Probabilistic models (histogram-based segmentation)
• Post-processing of results
• Model evaluation and reporting

The pipelines are designed to be modular and flexible, allowing experimentation with different features, models, and evaluation strategies.

Dataset

Land Cover for Aerial Imagery (Landcover AI)

• License: Creative Commons
• Kaggle: https://www.kaggle.com/datasets/aletbm/land-cover-from-aerial-imagery-landcover-ai
• Official website: https://landcover.ai.linuxpolska.com/

Minimal required data:

• images/ – satellite images
• labels/ – segmentation masks

Optional:

• raw_data/ – allows recreating Train/Validation/Test splits using the provided split.py script

Main Features

This project extracts multiple types of features from images for segmentation, grouped into main categories:

• Color and spectral: RGB, HSV, NDVI, Water Index
• Intensity and multi-scale context: Grayscale, Gaussian blur (multi-scale)
• Gradient and structural: Gradient magnitude & orientation, anisotropy, corner density
• Texture: Local variance, local entropy, LBP

Repository Structure

• src/ – All source code
• data/ – Dataset (images, labels, raw_data)
• logs/ – Project log files
• .trash/ – Old, useless or deleted files
• documents/ – Documentation files and plots

Inside documents/

• src_structure – Description of all source files
• data_structure – Structure of data folders
• FinalProject – Academic subject for the project
• plot/ – Plots for data analysis and evaluation
• evaluation_strategy – Explanation of evaluation methods

Installation

Recommended IDE: VS Code AI Code Assistance: GitHub Copilot, Claude Sonnet 4.5, ChatGPT 5.1 Python version: 3.10.19 Environment Manager: Conda

Create Virtual Environment

conda create -n IMG_SEG python=3.10
conda activate IMG_SEG

Install Dependencies

pip install -r requirements.txt

Hardware Requirements

• CPU: Intel i5-12400f (minimum)
• RAM: 16 GB DDR4
• GPU: NVIDIA RTX 4060 Ti (optional, recommended for faster training)

Training time estimates:

• CPU + downsampling, 50 epochs: ~6–8 hours
• GPU (50 epochs): depends on GPU, typically 20–40 minutes

Data Setup

Create the following folder structure in the data/ folder:

data/
├─ images/
├─ labels/
├─ raw_data/   (optional)

Use the provided split.py script to create Train/Validation/Test splits from the raw data.

If you encounter issues with paths, check the DataPath class in src/cste.py.

External Elements / Citations

• Landcover AI dataset - License: Creative Commons : CC BY-NC-SA 4.0
  • Kaggle: https://www.kaggle.com/datasets/aletbm/land-cover-from-aerial-imagery-landcover-ai
  • Official website: https://landcover.ai.linuxpolska.com/
• Autors of the dataset
  • Boguszewski Adrian
  • Batorski Dominik
  • Ziemba-Jankowska Natalia
  • Dziedzic Tomasz
  • Zambrzycka Anna
• Kaggle uploader pseudo
  • AleTBM

Contact / Support

For questions or issues regarding code execution:

Colin MANYRI – colin.manyri@etu.utc.fr