raphaelattias/overfitteam-geneva-satellite-images

Geneva Satellite Images Dataset - Rooftop Segmentation

Satellite Image	Segmentation Label

Dataset Description

This dataset contains high-resolution satellite imagery of Geneva, Switzerland, with corresponding segmentation labels for rooftop detection. The dataset was originally created for research on automated solar panel installation assessment using deep learning. It was developed as part of a study published in the Journal of Physics: Conference Series, focusing on quantifying suitable rooftop areas for solar panel installation using Convolutional Neural Networks. The dataset provides a comprehensive collection of aerial views with pixel-level annotations for building rooftops.

Dataset Summary

• Total Image-Label Pairs: 1,050
• Train Set: 840 image-label pairs (80%)
• Validation Set: 104 image-label pairs (10%)
• Test Set: 106 image-label pairs (10%)
• Image Format: PNG format
• Resolution: High-resolution satellite imagery tiles
• Categories: Three subsets (all, industrial, residential)
• Total Size: ~315 MB

Dataset Structure

The dataset is organized into train/val/test splits, with each split containing images and their corresponding segmentation labels organized by category:

overfitteam-geneva-satellite-images/
├── train/                 # 840 total image-label pairs
│   ├── images/
│   │   ├── all/           # 420 images - mixed building types
│   │   ├── industrial/    # 199 images - industrial buildings
│   │   └── residencial/   # 221 images - residential buildings
│   └── labels/
│       ├── all/           # 420 segmentation masks
│       ├── industrial/    # 199 segmentation masks
│       └── residencial/   # 221 segmentation masks
├── val/                   # 104 total image-label pairs
│   ├── images/
│   │   ├── all/           # 52 images - mixed building types
│   │   ├── industrial/    # 26 images - industrial buildings
│   │   └── residencial/   # 26 images - residential buildings
│   └── labels/
│       ├── all/           # 52 segmentation masks
│       ├── industrial/    # 26 segmentation masks
│       └── residencial/   # 26 segmentation masks
└── test/                  # 106 total image-label pairs
    ├── images/
    │   ├── all/           # 53 images - mixed building types
    │   ├── industrial/    # 28 images - industrial buildings
    │   └── residencial/   # 25 images - residential buildings
    └── labels/
        ├── all/           # 53 segmentation masks
        ├── industrial/    # 28 segmentation masks
        └── residencial/   # 25 segmentation masks

Features

Each sample in the dataset contains:

• Satellite Image: High-resolution PNG image tile from Geneva region
• Segmentation Label: Corresponding pixel-level mask for rooftop areas
• Category: Classification into all/industrial/residential building types
• Spatial Information: Encoded in filename (LV03 Swiss coordinates and year)

Filename Convention

Files follow the pattern: DOP25_LV03_{grid}_{year}_1_15_{x_coord}_{y_coord}.png

• Grid reference (e.g., 1301_11, 1301_13, 1301_31)
• Year (2015)
• Swiss LV03 coordinates (x, y)

Original Research Paper

This dataset was created for the following research:

Title: Quantification of the suitable rooftop area for solar panel installation from overhead imagery using Convolutional Neural Networks

Authors: Roberto Castello, Alina Walch, Raphaël Attias, Riccardo Cadei, Shasha Jiang, and Jean-Louis Scartezzini

Publication: Journal of Physics: Conference Series, Volume 2042, CISBAT 2021 Carbon-neutral cities - energy efficiency and renewables in the digital era

DOI: 10.1088/1742-6596/2042/1/012002

Key Results: The Convolutional Neural Network achieved an intersection over union of 64% and an accuracy of 93% for segmenting suitable rooftop areas for solar panel installation.

Use Cases

This dataset is suitable for:

• Rooftop Segmentation: Semantic segmentation of building rooftops from aerial imagery
• Solar Potential Assessment: Identifying and quantifying suitable rooftop areas for solar panel installation
• Building Detection: Training models to detect and segment buildings in satellite imagery
• Urban Analysis: Analyzing building density, rooftop characteristics, and urban morphology
• Segmentation Model Benchmarking: Evaluating U-Net, DeepLab, and other semantic segmentation architectures
• Transfer Learning: Pre-training on building segmentation for related remote sensing tasks

Loading the Dataset

Download from Hugging Face

from huggingface_hub import snapshot_download

# Download the entire dataset
dataset_path = snapshot_download(
    repo_id="raphaelattias/overfitteam-geneva-satellite-images",
    repo_type="dataset"
)

print(f"Dataset downloaded to: {dataset_path}")

Load Images and Labels

from PIL import Image
import os

# After downloading, load image and label
train_img_path = os.path.join(dataset_path, "train/images/all/DOP25_LV03_1301_11_2015_1_15_497500.0_119062.5.png")
train_label_path = os.path.join(dataset_path, "train/labels/all/DOP25_LV03_1301_11_2015_1_15_497500.0_119062.5_label.png")

image = Image.open(train_img_path)
label = Image.open(train_label_path)

PyTorch Dataset Class

from torch.utils.data import Dataset
import torchvision.transforms as transforms
from huggingface_hub import snapshot_download
from PIL import Image
import os

class GenevaRooftopDataset(Dataset):
    def __init__(self, split='train', category='all', transform=None, dataset_path=None):
        """
        Args:
            split (str): One of 'train', 'val', or 'test'
            category (str): One of 'all', 'industrial', or 'residencial'
            transform: Optional transform to be applied on images and labels
            dataset_path (str): Path to downloaded dataset. If None, will download automatically.
        """
        if dataset_path is None:
            dataset_path = snapshot_download(
                repo_id="raphaelattias/overfitteam-geneva-satellite-images",
                repo_type="dataset"
            )

        self.split = split
        self.category = category
        self.image_dir = os.path.join(dataset_path, f"{split}/images/{category}")
        self.label_dir = os.path.join(dataset_path, f"{split}/labels/{category}")
        self.transform = transform

        self.images = sorted([f for f in os.listdir(self.image_dir) if f.endswith('.png')])

    def __len__(self):
        return len(self.images)

    def __getitem__(self, idx):
        img_name = self.images[idx]
        label_name = img_name.replace('.png', '_label.png')

        image = Image.open(os.path.join(self.image_dir, img_name)).convert('RGB')
        label = Image.open(os.path.join(self.label_dir, label_name))

        if self.transform:
            image = self.transform(image)
            label = self.transform(label)

        return image, label

# Create dataset - will automatically download if needed
train_dataset = GenevaRooftopDataset(split='train', category='all')
val_dataset = GenevaRooftopDataset(split='val', category='all')
test_dataset = GenevaRooftopDataset(split='test', category='all')

print(f"Train samples: {len(train_dataset)}")
print(f"Val samples: {len(val_dataset)}")
print(f"Test samples: {len(test_dataset)}")

Data Collection

The satellite images were collected from aerial surveys of Geneva, Switzerland, focusing on:

• Residential Areas: Single-family homes, apartment buildings, and mixed residential zones with diverse roof types and orientations
• Industrial Zones: Warehouses, factories, and commercial buildings with large, flat rooftops ideal for solar installations
• Mixed Urban Areas: Downtown and suburban regions combining different building types and densities
• Geographic Coverage: Multiple grid sections covering different parts of Geneva (grid references 1301_11, 1301_13, 1301_31)
• Temporal Consistency: All imagery from 2015 for consistent lighting and seasonal conditions

The dataset enables machine learning models to learn features relevant to rooftop segmentation and solar panel suitability assessment, including roof shape, size, orientation, and surrounding context.

Label Information

The segmentation labels are binary masks where:

• Rooftop pixels: Marked areas suitable for solar panel installation
• Non-rooftop pixels: Background, vegetation, roads, other structures
• Labels are pixel-aligned with their corresponding images
• Masks are saved as PNG files with _label.png suffix

Preprocessing

The images and labels have been:

• Tiled into manageable patches for efficient training
• Georeferenced using Swiss LV03 coordinate system
• Quality-checked for cloud coverage and clarity
• Labeled with pixel-level precision for rooftop boundaries
• Split into train/val/test maintaining spatial distribution

Citation

If you use this dataset in your research, please cite both the dataset and the original paper:

Dataset Citation

@dataset{geneva_satellite_2024,
  title={Geneva Satellite Images Dataset - Rooftop Segmentation},
  author={Attias, Raphaël and Cadei, Riccardo and Jiang, Shasha},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/raphaelattias/overfitteam-geneva-satellite-images}
}

Original Paper Citation

@article{castello2021quantification,
  title={Quantification of the suitable rooftop area for solar panel installation from overhead imagery using Convolutional Neural Networks},
  author={Castello, Roberto and Walch, Alina and Attias, Raphaël and Cadei, Riccardo and Jiang, Shasha and Scartezzini, Jean-Louis},
  journal={Journal of Physics: Conference Series},
  volume={2042},
  number={1},
  pages={012002},
  year={2021},
  publisher={IOP Publishing},
  doi={10.1088/1742-6596/2042/1/012002}
}

License

This dataset is released under the MIT License, allowing for both academic and commercial use with attribution.

Acknowledgments

This dataset was created by the OverfitTeam, originally developed for research at EPFL's Solar Energy and Building Physics Laboratory (LESO-PB). The work was presented at CISBAT 2021 (Carbon-neutral cities - energy efficiency and renewables in the digital era) and published in the Journal of Physics: Conference Series.

Special thanks to:

• The research team at EPFL LESO-PB
• Co-authors of the original paper: Roberto Castello, Alina Walch, Riccardo Cadei, Shasha Jiang, and Jean-Louis Scartezzini
• The Hugging Face community for hosting and supporting this dataset

Contact

For questions, issues, or contributions, please open an issue on the dataset repository or contact the maintainers through the Hugging Face discussion forum.