OmniCloudMask

State-of-the-art cloud and cloud shadow segmentation for high to moderate resolution satellite imagery.

Works with any imagery containing Red, Green, and NIR bands at 10-50 m resolution (and down to 5 m with recent model versions). Validated on Sentinel-2, Landsat 8, PlanetScope and Maxar imagery. See choosing a resolution for guidance on picking the right resolution for your data.

Documentation | Paper | Training Data Map | Podcast

Installation

pip install omnicloudmask

Or with uv, conda, or from source—see installation docs.

Quick Start

import numpy as np
from omnicloudmask import predict_from_array

# Input: (3, height, width) array with Red, Green, NIR bands
input_array = np.random.rand(3, 1024, 1024).astype(np.float32)

# Output: (1, height, width) mask
# Values: 0=Clear, 1=Thick Cloud, 2=Thin Cloud, 3=Cloud Shadow
mask = predict_from_array(input_array)

For a Sentinel-2 scene:

from pathlib import Path
from omnicloudmask import predict_from_load_func, load_s2

scene_paths = [Path("path/to/scene.SAFE")]
pred_paths = predict_from_load_func(scene_paths, load_s2)

For a dataset of same-sized image chips, run several chips per batch:

from omnicloudmask import predict_from_batch

# chips: (N, 3, 512, 512) array, output: (N, 1, 512, 512) masks
masks = predict_from_batch(chips, batch_size=16)

See the quickstart guide for more examples.

Example Notebooks

• Sentinel-2 from .SAFE
• Sentinel-2 via Planetary Computer
• HLS (Harmonized Landsat Sentinel)
• PlanetScope
• PlanetScope Hyperspectral
• Maxar
• Batch prediction for image datasets

Try in Colab

How it works

License

MIT License