38-Cloud: Cloud Segmentation on Satellite Images
A satellite imagery dataset with pixelated annotations for cloud detection, based on 38 Landsat 8 scenes and over 78,000 patches.
78,900 files in total (384×384 patches), multi-channel images (R, G, B, NIR), mask annotations
Apache 2.0
Description
The dataset 38-Cloud offers a comprehensive database of Landsat 8 satellite images accompanied by accurate pixel-level annotations for cloud segmentation. It includes 38 satellite scenes divided into more than 78,000 384×384 images, with four spectral bands: Red, Green, Blue, and Near Infrared. This data is organized to facilitate use in convolutional neural networks dedicated to segmentation.
What is this dataset for?
- Train semantic segmentation models for cloud detection
- Evaluate cloud cover on satellite scenes in climate or agricultural projects
- Improving multispectral image processing algorithms
Can it be enriched or improved?
Yes, the dataset can be enriched by merging with other satellite data (weather, topography) or by artificially increasing the images (rotation, noise, zoom). It is also possible to refine the masks or adapt the spectral bands for other types of sensors.
🔎 In summary
🧠 Recommended for
- Satellite vision researchers
- Geospatial engineers
- Remote Sensing Projects
🔧 Compatible tools
- PyTorch
- TensorFlow
- Keras
- MMsegmentation
- OpenCV
💡 Tip
For better performance, combine the 4 spectral bands at the input of the models, rather than using only RGB.
Frequently Asked Questions
Can this dataset be used for weather forecasting projects?
Indirectly, yes. It makes it possible to analyze cloud cover, which can feed into weather or climate models.
Is it necessary to recompose the RGB strips manually?
Yes, each band is stored separately, they must be assembled for color rendering or coherent multispectral input.
Is it suitable for training a U-Net or a DeepLabV3+?
Absolutely, it is perfectly structured for modern semantic segmentation architectures like U-Net or DeepLab.
