We're sorry but this page doesn't work properly without JavaScript enabled. Please enable it to continue.
Feedback

Pan-european seasonal cloudless mosaic based on Sentinel-2 imagery

00:00

Formal Metadata

Title
Pan-european seasonal cloudless mosaic based on Sentinel-2 imagery
Title of Series
Number of Parts
57
Author
License
CC Attribution 3.0 Germany:
You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor.
Identifiers
Publisher
Release Date
Language
Producer
Production PlaceWageningen

Content Metadata

Subject Area
Genre
Abstract
Josip Križan is the owner of MultiOne. In this talk, Josip introduced how the rapidly increasing amount of publically available remotely sensed data in recent decades has revolutionized large-scale research and context-informed decision making. However, these data are generally not freely available as homogenized products ready for analysis at continental (or larger) scales. This is widely observed with datasets generated by EO satellites, particularly those with optical sensors and those capable of high-resolution imaging, where the process of mosaicking imagery to produce a homogenous, cloudless dataset across a particular area of interest often grows increasingly cumbersome at larger scales.
Keywords
Mathematical analysisProcess (computing)Scale (map)Product (business)SatelliteMedical imagingOpticsMusical ensembleSpring (hydrology)Source codeStatisticsAxonometric projectionImage resolutionMetreAddressing modeGreen's functionImage resolutionMedical imagingNear-ringProduct (business)Raw image formatProcess (computing)Source codeCountingScaling (geometry)Musical ensembleValidity (statistics)Projective planeComputing platformLevel (video gaming)Photographic mosaicOpen setDescriptive statisticsSpring (hydrology)PixelMetreSet (mathematics)Personal area networkArithmetic meanGoodness of fitDialectComputer animation
Musical ensembleSpring (hydrology)Computing platformZeno of EleaMaxima and minimaDaylight saving timeSource codeStatisticsMetreAddressing modeImage resolutionAxonometric projectionGreen's functionOrbitPhotographic mosaicComputer-generated imageryReading (process)AverageWeight functionPoint cloudAreaPixelAcoustic shadow19 (number)Bit rateRule of inferenceHypermediaGame theoryTesselationCountingOrbitValidity (statistics)Point cloudLink (knot theory)PixelAcoustic shadowNear-ringResultantSpacetimeMereologyPhotographic mosaicTheory of relativityMusical ensembleSphereUniform resource locatorPattern languageComputer animation
DemonLie groupEmbargoMusical ensembleOrbitTesselationComputer animation
OrbitMusical ensemblePhotographic mosaicComputer-generated imageryReading (process)Weight functionAverageGame theoryPixelComputer animation
Optical disc driveSimultaneous localization and mappingDifferent (Kate Ryan album)Centralizer and normalizerPixelProduct (business)Line (geometry)DistanceInverse elementWeightOrbitSatelliteMagnetic stripe cardMedical imagingInheritance (object-oriented programming)Level (video gaming)WhiteboardComputer animation
Solid geometryProduct (business)UsabilityGroup actionOrbitComputer animation
Rule of inferenceSelf-organizationMetreComputer animation
Execution unitInterior (topology)Musical ensembleSphereComputer animation
Total S.A.ScalabilityMusical ensembleComputer-generated imageryPhysical systemComponent-based software engineeringPhotographic mosaicMaxima and minimaParallel portServer (computing)Scripting languageClient (computing)OrbitTrailRead-only memoryMenu (computing)Information managementEmulationProcess (computing)Descriptive statisticsSelf-organizationComputer clusterInstance (computer science)Mobile appParallel portDatabaseAlgorithmBefehlsprozessorTrailServer (computing)Medical imagingMusical ensemblePhysical systemClient (computing)Scripting languageNumberMultiplicationTesselationOrbitPhotographic mosaicComputer animation
File viewerMultiplication signOpen setMoment (mathematics)Coefficient of determinationComputer animation
Execution unitMusical ensembleComputer-generated imageryPhysical systemComponent-based software engineeringPhotographic mosaicParallel portMaxima and minimaServer (computing)Scripting languageClient (computing)OrbitTrailRead-only memoryDuality (mathematics)Revision controlMetreSpring (hydrology)MechatronicsSummierbarkeitGamma functionService (economics)NumberCountingMathematical analysisDifferent (Kate Ryan album)Type theoryMultiplication signMoment (mathematics)Musical ensemblePixelNichtkommutative Jordan-Algebra1 (number)DataflowComputer animation
Transcript: English(auto-generated)
Well, I'm Josip Grisian. I come from Multivan company. So we made this pan European seasonal cloudless mosaic with Sentinel-2 images.
Can you see my screen? So there is a lot of public publicly available remote, remotely sensed data, and
it has been rapidly increasing in recent decades, but the data are not generally freely available as harmonized and analysis-ready products.
So the process of mosaicing these images to produce something that's homogeneous and cloudless can become a little bit cumbersome at larger scales.
So this is the description of data set. A source of data were Sentinel-2 level 2A open data on AWS. We produced, we took the six bands from there, blue, green, red, near infrared, CR1, CR2.
We took for every band, we took three statistics, 25, 50, 75 percent, and count of valid values, so per pixel.
We produced four seasons per year in 2018 and 2019, and winter and spring for 2020. So these are dates. These dates are aligned with GLAAD IRD products, so that it
was the idea that we made the fusion of this data set with the Landsat product. And all data is re-sampled to 30 meters resolution in this projection in 2035.
So data can be accessed in the Zenodo platform. You can download it from here.
Or there is also, we made the cloud-optimized geotiffs on Wasabi.
So it can be accessed directly via web. It can be accessed directly via through Pantungis.
This is the pattern of world links. And so for band, there is red, green, blue, near sphere one, sphere two. And for seasons, you have to replace this start date and end date in the links.
If you want to access this with the GLAAD tools, then you have to add prefix ZI curl example. So this is example. How can you access this data?
And yes, this data are, yes, and this data are scaled to bytes. Originally, data is in 16 bits, but just for a matter of space of size.
Size is very big, so you have to scale it to bytes. So this is how it looks. I can show you, I will show you in Pantungis.
For example, this is summer 2019, and this is autumn 2019. It's not perfect, especially this northern part.
So how it's done? First, we calculate the mosaic of one season, one band in one centimeter of tile and relative orbit.
Just so all the big values that has clouds or cloud shadows is masked out, and then it's calculated 25, 50, 75 percentile and count for valid values for every pixel location.
And this intermediate result is saved back to temporary S3 location. So this is done for all, there was 137 centimeter tiles.
So this is a grid of centimeter tiles. So first it was every tile is calculated independently.
And then all tiles that are in the same season, band and orbit is stitched together. It's just taking mean value of overlapping, well not tiles, overlapping pixels.
So this looks, it looks like something like this. So we have these tiles, but we have these orbits. So everything that is in one orbit is consistent because a satellite is, all these images are taken one after another.
So there is no visible difference between them. But when we put the neighboring orbit one by another,
then we have, see this, that we have, they are not aligned perfectly. So, so this was, these two orbits are then stitched together by using inverse distance weight mean
with respect to the distances, distances from all, from every pixel to, to this central orbit line.
So that way, when we put this all orbits together, we can get, we have this final product where there is no stripes.
There is no visible, there is no visible transitions from one orbit to another. Yes, this is, so for example, winter 219, RGB, RGB composite looks something like this.
There is, there is a lot of this white values because there is just, it's just snow. Snow isn't, wasn't masked out.
And summer.
Maybe you could zoom into some areas. Yes, I can zoom it. So this is directly from Wasabi. It's really fast.
So this is, oh, that's it. It's the 30, 30 meters resolution. And okay, this is RGB composite. I made this near-sphere composite too. It looks interesting.
This is, so this is from near-sphere one and sphere two bands. Okay. So yes, that was the algorithm that we used.
And this was computed on multiple spot instances on the AWS. Every instance have 64 to 72 with CPUs and 500 to 700 gigabytes of RAM. And so this one, so every season band tile orbit mosaic was one independent job.
And it's a, it could be populated in a parallel. So there is a system is made of two components.
There is client script that runs on this multiple instances and it fetches job descriptions from an central server. And that server, there is an app behind this app is a Postgres database that keeps track of all jobs done, all jobs that need to be done.
And take care that there is no duplicate jobs and so on. So this approach is scalable to any number of instances with CPUs. So of course, up to total number of jobs. It's not, well, it's not really correct this scalable.
If you have too many instances, then this central server becomes quite too many, but okay, that can be. So there is 137 S2 tiles, and it was around 42 to 43,000 images, Sentinel 2 images for each of the band.
So for each season to make this all mosaics, it was really a lot of data.
Well, that's, well, that's it. I was really quick. Okay. Any questions?
Okay, thank you. And so this data, can you go back to Zenodo? Just show us one more time at the moment. So it's available in Zenodo and it will be put also in the open data viewer. And also at the moment, it's available through our service. We don't have to download everything. But it will, basically the data is ready.
What about the gap filling? Well, there's no, there's no, we didn't make any gap filling. We have these count layers where there is, for every pixel, there is a count of numbers.
So if it's low number, then this pixel is not really reliable. So it can be used like a gap filling, but we didn't make any gap filling at the moment for this. Okay, but it's a fantastic piece of work. And we have both Landsat and Sentinel.
We have inside the project, the prepared analysis ready. Okay, with your data, the gaps with the Landsat, we fill in all the gaps. What is your impression of the difference between Landsat and Sentinel? Because there's a, you know, it's the same resolution, it's the kind of multispectral, same type of bands.
What is your impression so far about the differences Landsat and Sentinel? Well, Landsat has 16 days with revisiting time. Sentinel has five, five years. So it's almost three times more data on the timeline.
So that's the big deal. So there is less, there is less gaps for that. But then Landsat has, this Landsat 7, Landsat 5, Landsat 8.
So that's you have, you have more than 20 years of Landsat. While Sentinel is only from 200 and 16, 17. So,