Mount Taranaki cuts through the clouds. Sentinel-2 image from 2017–12–15 via Sentinel Hub with a cloud mask produced with the Sentinel Hub Cloud Detector at three different zoom levels.

Improving Cloud Detection with Machine Learning

Sentinel-2 image of Betsiboka Estuary recorded on 2017-12-15 via Sentinel Hub without and with a cloud mask produced with the Sentinel Hub Cloud Detector. The algorithm correctly identifies regions of thin and transparent clouds. Move the slider for a better visual inspection.
Cloud masking of Sentinel-2 images in real time using Sentinel Hub services and Sentinel Hub Cloud Detector.

Cloud detection algorithms in a nutshell

Sentinel-2 image of snow covered Etna recorded on 2017-12-14 via Sentinel Hub without and with the Sentinel Hub Cloud Detector.

Sentinel Hub’s pixel-based cloud detector

Sentinel-2 image of Sydney recorded on 2017-12-06 via Sentinel Hub without and with the Sentinel Hub Cloud Detector.

Training and validation samples

Location of 108 Sentinel-2 scenes used by Hollstein et al. for manual classification. 15 scenes indicated with yellow marker represent scenes that are processed by MAJA and available for download. We use these to validate MAJA.
Distribution of 596 tile locations from which we randomly sample pixels to build our training set.
Sentinel-2 image of Maui recorded on 2017-12-12 via Sentinel Hub without and with the Sentinel Hub Cloud Detector.

Model selection and hyper-parameter tuning

Training set augmentation

Sentinel-2 image of clouds over water (left) and corresponding cloud masks as determined by MAJA (middle) and earlier versions of Sentinel Hub Cloud Detector (right). Clouds are black.
Sentinel-2 image of bare land with a small cloud in lover right area (left) and corresponding cloud masks as determined by MAJA (middle) and earlier versions of Sentinel Hub Cloud Detector (right). Clouds are black.

Validation and comparison with other cloud detectors

Cloud and cirrus cloud detection rates and land, water, snow and shadow misclassification rates as clouds as determined using 108 Sentinel-2 scenes hand labeled by Hollstein et al.
Cloud and cirrus cloud detection rates and land, water, snow and shadow misclassification rates as clouds as determined using 15 Sentinel-2 scenes hand labeled by Hollstein et al.
Sentinel-2 image of Abarrancamento recorded on 2017-12-15 via Sentinel Hub without and with the Sentinel Hub Cloud Detector.

Concluding remarks

Interested in using Sentinel Hub Cloud Detector?

Sentinel Hub Blog

Stories from the next generation satellite imagery platform

Anze Zupanc

Written by

Physicist | Data Scientist | Problems Solver

Sentinel Hub Blog

Stories from the next generation satellite imagery platform