Solar Panel Segmentation from Aerial Images ☀️

CS-433 | Machine Learning @ EPFL

Mehdi Zoghlami, Maxime Ducourau, Jean Perbet

Introduction

This repository contains the code for the second project of the course CS-433 - Machine Learning @ EPFL. This project was run in partnership with the Swiss Data Science Center, and the goal was to develop a machine learning model to detect solar panels from aerial images provided by Swisstopo.

This project experiments with both U-Net and DeepLabv3 architectures, using CNNs for the binary semantic segmentation task of deciding whether a pixel is a solar panel or not on aerial images.

Repository structure

The repository is structured as follows:

├── data/                           
│   ├── images/              # Images
│   ├── labels/              # Ground truth labels
│   ├── roofs/                        
│   │   ├── coordinates/     # Coordinates of all swiss roofs
│   │   ├── images/          # Images w/ filtered out roofs
│   │   └── masks/           # Binary masks of roofs corresponding to images
│   ├── tiles/               # Original tiles             
│   ├── urls/                # Swisstopo URLs  
│   └── weights/             # Pre-trained weights for models
├── notebooks/
│   ├── preprocessing.ipynb  # Notebook for data retrieval & preprocessing
│   └── results.ipynb        # Main notebook w/ results
├── report.pdf               # Associated paper
├── src/
│   ├── models/              # Model-specific classes/functions    
│   ├── dataset.py           # Data loading classes/functions
│   ├── model_functions.py   # Generic model functions
│   ├── plotting.py          # Plotting functions
│   ├── postprocessing.py    # Postprocessing functions
│   └── preprocessing.py     # Preprocessing functions
└── README.md

Run instructions

1. Install the required dependencies.

   pip install -r requirements.txt

2. Data should be collected and put in the corresponding directories. There are two ways to do this.

   • Run the preprocessing.ipynb notebook in the notebooks/ directory (~500 data points)
   • Directly download the data from this kaggle dataset (~750 data points)

3. Decide whether you want to train models on your own or use them pre-trained with the weights we provide. In the latter case, you should download the weights from the same kaggle dataset as for the previous step.

4. Run the main notebook results.ipynb with training & results parts.