Sentinel-2 Super-resolution takes a Sentinel-2 image and applies deep learning techniques to improve the resolution of the multi-spectral bands. A Sentinel-2 image has 13 spectral bands which are obtained at three different spatial resolutions (10m, 20m, and 60m). In order to achieve the higher spatial resolution for all the spectral bands within 20m and 60m, the super-resolution algorithm utilizes a convolutional neural networks (CNN) to perform the upsampling. Training of these networks are based on lower spatial resolutions which are obtained by downsampling the 20m and 60m. As mentioned by the author, these networks are trained based on images from different climate zones and land-cover types to achieve a good generalization. Therefore, this block employs the pre-trained networks and returns super-resolved images with the GeoTIFF format. The modified version of the source code is published as Free Software / Open Source on GitHub.
Please note: Block Version 2.0.6-public supports clipping scenes. Set "clip_to_aoi" to True and use the same coordinates for the Superresolution parameters in Bounding Box, Contains or Intersect. The pricing for this block is based on the megabyte output size.
Minimum AOI: The minimum required AOI size is 4 sqkm. Jobs with lower AOI sizes fail when
clip_to_aoi is set to "true".
|Block Type||Processing (data preparation)|
|Supported Input Types||Sentinel 2 L1C GRD in SAFE Format|
|Resolution||10 m for all spectral bands|
|Performance||Delivers output with highest spatial resolution. GPU is recommended as CPU for the whole image may be computationally expensive.|
Click here to learn more about the block's documentation.
[ "coastal", "blue", "green", "red", "rededge", "rededge2", "rededge3", "nir", "nir2", "watervapour", "swir", "swir2", "swir3" ]
End User License Agreement
View the End User License Agreement conditions in a new tab.