Spectral homogenization of multispectral satellite data.
SpecHomo - Spectral homogenization of multispectral satellite data
- Free software: GNU General Public License v3 or later (GPLv3+)
- Documentation: https://geomultisens.gitext-pages.gfz-potsdam.de/spechomo/doc/
- The paper corresponding to this software repository can be found here: Scheffler et al. 2020 (cite as: Scheffler D., Frantz D., Segl K. (2020). Spectral harmonization and red edge prediction of Landsat-8 to Sentinel-2 using land cover optimized multivariate regressors. Remote Sens. Environ. 241, 111723. https://doi.org/10.1016/j.rse.2020.111723)
- Information on how to cite the SpecHomo Python package can be found in the CITATION file.
- Submit feedback by filing an issue here or join our chat here:
SpecHomo is a Python package for spectral homogenization of multispectral satellite data, i.e., for the transformation of the spectral information of one sensor into the spectral domain of another one. This simplifies workflows, increases the reliability of subsequently derived multi-sensor products and may also enable the generation of new products that are not possible with the initial spectral definition.
SpecHomo offers different machine learning techniques for the prediction of the target sensor spectral information. So far, multivariate linear regression, multivariate quadratic regression and random forest regression are implemented. To allow easy comparisons to the most simple homogenization approach, we also implemented linear spectral interpolation.
In contrast to previous spectral homogenization techniques, SpecHomo not only allows to apply a global (band-wise) transformation with the same prediction coefficients for all gray values of a spectral band. It also distinguishes between individual spectral characteristics of different land-cover types by using specifically trained prediction coefficients for various spectral clusters. This increases the accuracy of the predicted spectral information. Apart from that, SpecHomo can not only be used to homogenize already similar spectral definitions - it also allows to predict unilaterally missing bands such as the red edge bands that are not present in Landsat-8 data.
Prediction accuracies and effects to subsequent products such as spectral indices or classifications have been evaluated in the above mentioned paper at the example of Sentinel-2 spectral information predicted from Landsat-8. Algorithm details may also be found there.
Satellite data (surface reflectance) acquired by following sensors may be used as source or target sensor:
- Landsat-5 TM
- Landsat-7 ETM+
- Landsat-8 OLI
- Sentinel-2A MSI
- Sentinel-2B MSI
- RapidEye-5 MSI
SpecHomo features classifiers for homogenization that we trained in the context of the GeoMultiSens project (see the credits section) and for our evaluations related with the above mentioned paper. The initial spectral information for classifier training has been derived from hyperspectral airborne data, spectrally convolved to different sensors. You may also train your own homogenization classifiers specifically optimized to your area of interest. SpecHomo provides the needed functionality for that.
For further details on how to use SpecHomo check out the documentation!
- Fixed incorrect handling of classification map nodata value in SpectralHomogenizer (default is now -9999).
- Updated minimal version of specclassify to 0.2.8. This fixes issue #8 (Homogenization using kNN classifiers uses faulty weights in case SpectralHomogenizer.predict_by_machine_learner() is called with global_clf_threshold=None.).
- Changed minimum and maximum values used for normalization of spectral distance measures. SAM values are now normalized between 0 and 15 degrees SA and all other measures use the 90% percentile as maximum value. This fixes issue #9 (Harmonization weights are affected by extreme values in the computed spectral distances between spectrum and available regressors.).
- Cluster classifiers can now be saved as JSON files (relates to issue #5).
- Added attributes ‘spechomo_version’ and ‘spechomo_versionalias’ to ClusterLearner.
- Added URL checker and corresponding CI job.
- Removed travis related files.
- Replaced hard-coded links in documentation by cross-linking directives.
- Replaced deprecated ‘source activate’ by ‘conda activate.’
- Updated installation instructions.
- Revised requirements.
- Added doc, test, lint and dev requirements to optional requirements in setup.py.
- Updated LR and QR classifiers.
- Added sklearn import to avoid static TLS ImportError.
- Improved code style of SpectralHomogenizer.interpolate_cube() and SpectralHomogenizer.predict().
- Bugfix for also predicting spectral information for pixels that contain nodata in any band (causes faulty predictions).
- Bugfix for only choosing 25 spectra in classifier creation in case the maximum angle threshold is automatically set to 0 because there are many well matching spectra.
- Added minimal version of geoarray.
- Use SPDX license identifier and set all files to GLP3+ to be consistent with license headers in the source files.
- Added latest QR classifiers.
- SpecHomo is now on conda-forge! Updated the installation instructions accordingly.
- Re-trained LR classifiers.
- Updated classifiers within test data.
- Classifiers are no longer stored in the repository (resources directory) but are automatically downloaded on demand at the first run (added corresponding code).
- Fixed TemporaryDirectory bug in Test_Utils.test_export_classifiers_as_JSON().
- Re-enabled CI job ‘deploy_pypi’.
- Fixed an AssertionError within ClusterClassifier_Generator.create_classifiers() caused by nodata pixels in the target sensor reference cube that were not dropped before creating the classifier.
- Moved matplotlib imports function/class level to avoid static TLS ImportError.
- Moved scipy imports function/class level to avoid static TLS ImportError.
- environment_spechomo.yml now uses Python 3.7+.
- scikit-learn is now pinned to 0.23.2+ due to classifier recreation.
- Fixed a DeprecationWarning in case of scikit-learn>=0.23.
- Dumped regressors now use the second highest dill protocol in order to have some downwards compatibility.
- Moved imports of scikit-learn to function/class level to avoid static TLS ImportError.
- Replaced deprecated HTTP links.
- Fixed test_spechomo_install CI job.
- Fixed create_github_release CI job.
- Releases in the GitHub-Mirror-Repository are now created automatically (added create_release_from_gitlab_ci.sh and create_github_release CI job).
- Added GitHub issue template.
- Revised CITATION file and .zenodo.json.
- Added functionality to export existing .dill classifiers to JSON format to make them also usable in different programming environments.
- The documentation now contains links to the published version of the research paper corresponding to SpecHomo.
- Changed Zenodo title and description.
- Fixed fallback algorithm in SpectralHomogenizer.predict_by_machine_learner() and added corresponding tests.
- SpectralHomogenizer.interpolate_cube() now returns a GeoArray instead of a numpy array.
- Removed pyresample dependency (not needed anymore).
- Updated README.rst and setup.py.
- Pinned geopandas to below version 0.6.3 to fix an incompatibility with pyproj.
- Updated CI runner setup scripts and CI jobs.
- Updated LR and QR classifiers.
- Added Sphinx documentation.
- Improved usability by adding functions to explore available spectral tansformations.
- All tests are working properly now.
- Added license texts.
- Revised global classifiers.
- Added harmonization using weighted averaging.
- A lot of algorithm improvements. Refer to the commits for details.
- First version working separately from geomultisens.
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size spechomo-0.9.3.tar.gz (452.4 kB)||File type Source||Python version None||Upload date||Hashes View|