The s2cloudmask Python package provides machine learning classifiers for Cloud and Shadow Detection in Sentinel-2 observations. The aim of this package is to open-source and showcase some of the tools being developed as part of the Digital Earth Australia initiative, and further, to push the state-of-art in the area of cloud classification.
The package currently provides a number of classifiers:
spectral: A spectral pixel-based cloud classifierfast: A spectral pixel-based cloud classifier (using a decision tree for speed and interpretability)temporal: A spectral-temporal pixel-based cloud classifierfast-shadow: A spectral-temporal pixel-based shadow classifier (using a decision tree for speed and interpretability)
The spectral classifiers are useful if you only have a couple of observations (i.e., satellite images) while the the spectral-temporal classifiers (aka. temporal classifiers) give better classifications of clouds (and shadows) provided that you can supply it with a geomedian pixel-composite mosaic [Roberts et al. 2017] of the area (or a stack of data so that one can be created by this package).
We note the existence of Python package s2cloudless developed by Sentinel Hub's research team that, as they argue in their blog post, "didn't observe significant improvement using derived features instead of raw band values" so their "final classifier uses the following 10 bands as input: B01, B02, B04, B05, B08, B8A, B09, B10, B11, B12". By releasing this package, we argue the contrary and demonstrate that you can obtain a better classification of clouds by (thinking hard and) developing new derived features for your machine learning algorithm.
In the image above: Baseline is s2cloudless, Spectral is our spectral classifier, Temporal is our temporal classifier.
$ pip install git+https://github.com/daleroberts/s2cloudmask
This package has an easy interface. Given a numpy array obs ordered as (y,x,band) we can obtain a cloud mask.
>>> import s2cloudmask as s2cm
>>> mask = s2cm.cloud_mask(obs, model='spectral')
Tests (and examples) are available in tests/test_.py and can be run with pytest from the project root.
$ pytest
You may be interested to read:
Roberts, D., Mueller, N., McIntyre, A. (2017). High-dimensional pixel composites from Earth observation time series. IEEE Transactions on Geoscience and Remote Sensing, PP, 99. pp. 1--11.
or maybe some of my other open-source projects.
