Table of Contents:
- rsdtlib: Remote Sensing with Deep-Temporal Data Library
- Documentation
- Example
- Use rsdtlib for ERCNN-DRS
- Use rsdtlib for monitoring Mariupol, Ukraine
- Paper and Citation
- Contact
- Acknowledgments
- License
This project provides a Python module that allows:
- Stage 1: Download remote sensing data directly from Sentinel Hub (i.e. Sentinel 1 & 2), or convert existing
GeoTIFFfiles - Stage 2: Temporally stack, assemble, and tile these observations
- Stage 3: Create windows of longer time series comprising these observations (i.e. deep-temporal)
Below figure shows the processing pipeline considering all three stages:
The processing in stages 2 and 3 is detailed below, showing how observations are combined to produce the final windows:
The documentation is hosted on Github Pages here.
All examples are in folder examples. These examples, as shown below, demonstrate each stage.
Note: Except for the download stage, all examples can be executed directly. For the download stage, the credentials for Sentinel Hub need to be added to the script to work.
Observations can either be directly downloaded form Sentinel Hub or pre-processed locally as GeoTIFF files and converted to EOPatch samples.
The following downloads all Sentinel 1 and 2 observations from first half of 2017, for the region specified in a shape file (here: Ostrava/CZ).
$ python example_download.py
Data progress: 100.0%
Downloaded to ./obs/S1_asc: 60
Data progress: 100.0%
Downloaded to ./obs/S1_dsc: 60
Data progress: 100.0%
CLM progress: 100.0%
100%||||||||||||||||||||||||||||| 36/36 [00:01<00:00, 31.61it/s]
Downloaded to ./obs/S2: 36
If preprocessed GeoTIFF files are already available, these can be easily converted to EOPatch samples, for the region specified in a shape file (here: Ostrava/CZ). In the example below, a single Level 1 Landsat 5 TM observation from Octrober 1993 is converted. Note that the GeoTIFF covers a larger area but the conversion only considers (subsets) the region from the shape file.
$ python example_convert.py
EOPatch(
data={
Bands: numpy.ndarray(shape=(1, 46, 87, 7), dtype=float32)
}
mask={
Mask: numpy.ndarray(shape=(1, 46, 87, 1), dtype=uint16)
}
bbox=BBox(((18.14081078, 49.83455151), (18.17269592, 49.85132019)), crs=CRS('4326'))
timestamp=[datetime.datetime(1993, 10, 10, 9, 0, 51)]
)
$ python example_stack.py
Resolutions (y, x):
OPT: 178, 236
SAR: 178, 236
Total time stamps: 138
Effective time stamps (from 'starttime' with 'delta_step' steps): 65
List of observations to process:
2017-01-01 05:00:30
2017-01-01 10:04:07
...
2017-06-30 05:00:20
2017-06-30 10:00:25
Total writes: 65
The following creates windows with labels (just dummies with values of one) and stores them on the file system.
$ python example_window_training.py
List of window ranges (current):
2017-01-02 04:52:47 2017-02-01 04:52:17
2017-01-04 16:34:04 2017-02-01 04:52:17
...
2017-06-01 04:52:06 2017-06-30 10:00:25
2017-06-03 16:34:37 2017-06-30 10:00:25
Writing training files:
Progress: 90.0%
Writing validation files:
Progress: 75.0%
For inference, either the windows can be stored on the file system (offline) or directly used with a model (online). Note that labels are not created in any of the inference examples, but is possible, too.
$ python example_window_inference.py
List of window ranges (current):
2017-01-02 04:52:47 2017-02-01 04:52:17
2017-01-04 16:34:04 2017-02-01 04:52:17
...
2017-06-01 04:52:06 2017-06-30 10:00:25
2017-06-03 16:34:37 2017-06-30 10:00:25
Progress: 97.1%
$ python example_window_interactive_inference.py
List of window ranges (current):
2017-01-02 04:52:47 2017-02-01 04:52:17
2017-01-04 16:34:04 2017-02-01 04:52:17
...
2017-06-01 04:52:06 2017-06-30 10:00:25
2017-06-03 16:34:37 2017-06-30 10:00:25
Progress: 97.1%
In our previous work of ERCNN-DRS Urban Change Monitoring the time series processing of (an early version of) rsdtlib was used. See subdirectory ./ERCNN-DRS for the pre-processing scripts.
The rsdtlib library was also used in our work of Monitoring Urban Changes in Mariupol/Ukraine. See subdirectory ./urban_change_monitoring_mariupol_ua for the pre-processing scripts.
The full paper can be found at SoftwareX.
@Article{ZITZLSBERGER2023101369,
AUTHOR = {Georg Zitzlsberger and Michal Podhoranyi and Jan Martinovič},
TITLE = {rsdtlib: Remote sensing with deep-temporal data library},
JOURNAL = {SoftwareX},
VOLUME = {22},
YEAR = {2023},
PAGES = {101369},
URL = {https://www.sciencedirect.com/science/article/pii/S2352711023000651},
ISSN = {2352-7110},
DOI = {https://doi.org/10.1016/j.softx.2023.101369}
}
Should you have any feedback or questions, please contact the main author: Georg Zitzlsberger (georg.zitzlsberger(a)vsb.cz).
This research was funded by ESA via the Blockchain ENabled DEep Learning for Space Data (BLENDED) project (SpaceApps Subcontract No. 4000129481/19/I-IT4I) and by the Ministry of Education, Youth and Sports from the National Programme of Sustainability (NPS II) project “IT4Innovations excellence in science - LQ1602” and by the IT4Innovations Infrastructure, which is supported by the Ministry of Education, Youth and Sports of the Czech Republic through the e-INFRA CZ (ID:90140), and via the Open Access Grant Competition (OPEN-21-31 and OPEN-25-24).
The authors would like to thank ESA for funding the study as part of the BLENDED project1 and IT4Innovations for funding the compute resources via the Open Access Grant Competition (OPEN-21-31 and OPEN-25-24). Furthermore, the authors would like to thank the data providers (USGS, ESA, Sentinel Hub and Google) for making remote sensing data freely available:
- Landsat 5 TM courtesy of the U.S. Geological Survey.
- ERS-1/2 data provided by the European Space Agency.
- Contains modified Copernicus Sentinel data 2017-2021 processed by Sentinel Hub (Sentinel 1 & 2).
The authors would finally like to thank the BLENDED project partners for supporting our work as a case study of the developed platform.
This project is made available under the GNU General Public License, version 3 (GPLv3).