Hugging Face's logo

Datasets:
FederCO23
/
solar-plants-brazil

Tasks:
Image Segmentation
Object Detection
Modalities:
Image
Geospatial
Formats:
imagefolder
Size:
< 1K
Tags:
satellite-imagery
remote-sensing
semantic-segmentation
photovoltaic
torchgeo
Libraries:
Datasets
License:
Dataset card Data Studio Files Community
1
solar-plants-brazil / README.md
FederCO23's picture
FederCO23
Create README.md
995e426 verified
preview code
|
raw
history blame
3.64 kB
metadata
license: mit
task_categories:
  - image-segmentation
  - object-detection
tags:
  - satellite-imagery
  - remote-sensing
  - semantic-segmentation
  - photovoltaic
  - torchgeo
pretty_name: Solar Plants Brazil (semantic segmentation)
size_categories:
  - n<1K

🛰️ Solar Plants Brazil

Solar Plants Brazil is a geospatial dataset for binary semantic segmentation of photovoltaic (PV) solar power stations in satellite imagery. It consists of multi-spectral image tiles (including near-infrared) with pixel-level annotations indicating the presence of solar panels. This dataset enables training and evaluating deep learning models that automatically detect solar farm installations from overhead imagery, supporting applications in renewable energy mapping and remote sensing.

Dataset Overview

  • Purpose: Identify photovoltaic solar plant locations via binary segmentation (panel vs. background) in satellite images
  • Imagery: 4-band (RGB + Near-Infrared) GeoTIFF tiles at 256×256 pixels each (float32 pixel values)
  • Labels: Corresponding binary mask for each image (1 = solar panel, 0 = no panel)
  • Volume: 272 image/mask pairs in total, split into training, validation, and test sets (201 train, 35 val, 36 test)
  • Geography: Tiles are from various locations across Brazil, focusing on known solar farm sites
  • Metadata: GeoTIFF files include spatial reference (CRS) and geolocation

📁 Dataset Structure 

solar-plants-brazil/
├── train/
│   ├── input/       # 256x256 RGBNIR images (GeoTIFF)
│   └── labels/      # 256x256 binary masks (GeoTIFF)
├── val/
│   ├── input/ 
│   └── labels/
└── test/
    ├── input/
    └── labels/

Each image is named img(<ID>).tif with its corresponding mask target(<ID>).tif. All images are pre-cropped to 256×256 pixels. The masks are binary, with 1 representing solar panels.

🔧 TorchGeo Integration

This dataset is integrated with the TorchGeo library and can be loaded as:

from torchgeo.datasets import SolarPlantsBrazil

dataset = SolarPlantsBrazil(split="train", root="data/", download=True)
sample = dataset[0]
image_tensor, mask_tensor = sample["image"], sample["mask"]
print(image_tensor.shape, mask_tensor.shape)

This returns image tensors with 4 channels and single-channel binary masks.

💡 Use Cases

  • Training segmentation models for solar farm detection
  • Monitoring solar plant growth over time
  • Infrastructure planning and spatial analytics
  • Academic research on geospatial ML

📚 Related Resources

📜 Citation 

@misc{solarplantsbrazil2024,
  author       = {Federico Bessi},
  title        = {Solar Plants Brazil: A Semantic Segmentation Dataset for Photovoltaic Panel Detection},
  year         = {2024},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/FederCO23/solar-plants-brazil},
  note         = {Preprocessing and training code available at https://github.com/FederCO23/UCSD_MLBootcamp_Capstone}
}

🙋‍♂️ Contact