The Weather Problem
Sentinel-2 is an optical satellite. It relies on sunlight reflecting off the Earth’s surface to capture an image. This means it has a massive, unavoidable blind spot: clouds.
If a thick cloud bank is sitting over Gaza on the day of a 5-day revisit, the satellite simply captures a picture of the top of the clouds. Furthermore, even if the sky is mostly clear, clouds cast long, dark shadows that completely obscure the ground data beneath them.
For a platform like Sentinel Bird, which relies on building continuous, high-fidelity timelapses and accurate change-detection models, contaminated pixels are worse than useless—they are actively misleading. A cloud shadow can look mathematically similar to a destroyed building or a burned field if you aren’t careful.
The Smoke Problem
In a warzone, the weather isn’t the only thing blocking the view. Massive plumes of thick, black smoke from explosions and fires frequently blanket the sky.
To a basic satellite algorithm, a thick cloud of smoke looks almost identical to a thick cloud of water vapor. If we use rudimentary filtering, we risk throwing out the exact imagery that captures the immediate aftermath of a bombardment.
The OmniCloudMask Solution
To solve this, Sentinel Bird integrates OmniCloudMask, a state-of-the-art deep learning ensemble trained on the CloudSEN12 dataset.
Instead of relying on simple brightness thresholds, OmniCloudMask processes three spectral bands (Red, Green, Near-Infrared) simultaneously to understand the texture and context of the noise. It classifies every single pixel in the image into distinct categories:
- Clear: Usable ground data.
- Cloud: Thick water vapor/ice blocking the view.
- Shadow: The dark cast of a cloud obscuring the ground.
The model then applies morphological post-processing to remove noise and smooth the boundaries of the masks. By automatically stripping out the contaminated pixels, we ensure that the interactive sliders and timelapses you see on this site are built exclusively from clean, reliable data.
What’s Next?
With clean, cloud-free pixels and a deep understanding of the spectral bands, we can finally start doing the math. In the final section, we will explore how to combine these bands to create indices that highlight vegetation, water, and destroyed infrastructure.