So as was already said, my name is Michal Kaczmorek, I'm working at the Department of Geoinformatics at the Technical University of Ostrava in the Czech Republic.

Today I will present work from a whole team. So I'm just one member, one piece of Mosaic. And it will be work related to finished project, which we were working on in a couple of last years. It was a project funded by European Space Agency. Its main contractor was a company CGI, IIT Czech Republic.

And the main responsibilities of the university of our team was to develop automatic processing change for Sentinel-1 imagery for specified use cases.

Two of them were based on intensity amplitude change detection, and the other ones were dealing with interferometry technique. And my main goal of this talk will be to introduce you the solutions which

we developed for individual use cases to say use something about their performance and so on. So let's start. For those of you who are not really familiar with Sentinel-1 mission, there is a pair of satellites with synthetic aperture radar, Sentinel-1A, Sentinel-1B, one of them is unfortunately not working since last Christmas.

All the imagery is of course publicly available, typically a few hours after its acquisition within the Czech Republic, each area can be set to be acquired from three or four independent tracks.

The spatial resolution of the data is dependent on the used product in our case, or in the case of our solutions, it's five or 20 meters.

So, not to repeat the given information for every solution which will be introduced, I will say something common to them. All of them are based on open source software, mainly OST tool and SNAP software developed by ESA, and also they are made from our own Python 3 scripts.

Most of the solution need just the Sentinel-1A imagery and information about the array of interest to be run. All of them can be run in two modes. The first one is like a simple one, you choose array of interest, time period and just run the solution.

The other one is meant for automatic continuous monitoring over the given array of interest, so anytime there is a new image is being downloaded, included in the processing process and the result is provided to the user.

So, I will start with the introduction of the first use case, which in the beginning of the project was more general like targeting on identification of any kind of change in a zone of, in a protected area of critical infrastructure so simply speaking in areas

where any kind of construction, a lot of works is just prohibited by law. During the project, there was, or the use case was mainly focusing on trying to detect a new building construction just from the Sentinel-1A imagery.

And although you can find in literature some change detection algorithms based on SAR data, which are aiming on urbanization, only few of them really were dealing with detection of individual single buildings.

And all of these were focusing, or were working with very high resolution data. So, our solution is based on a statistical approach. So we have a stack of images, pre-event one, and the post-event one, so you create images over some time period before the construction or any kind of work in which you are interested to detect.

And you statistically test if the difference between mean of backscatter intensity and coherence

in the pre-event period and post-event period has changed on a pixel level. Afterwards, there are some post-processing steps being done, including fuzzy technique region growing and so on, mainly to suppress false positive detections.

So, unfortunately, we do not have time to go deeply into description of the solution, but we will at least stop for a moment to say something about the performance of the developed solution. So, for the validation, we had two testing areas, one in an urban area, one in a mainly rural area with natural surfaces.

We have processed with the solution several years of data for each change which was detected, we were trying to classify on which type of surface it was detected.

And we were also, of course, assessing the success rates of the real building detections. So, providing you a summary of these results, we can say that when you process just data from one single track, you will be able to detect only about 40% of new buildings.

The success rates will be higher for buildings which are larger than standard single family house. And it leads to the result that the spatial resolution of the Sentinel-1 is the main limitation of the detection success. So, if we were using with the solution data with high resolution, we probably would get two better results.

A good result was that almost all of the changes were detected on artificial surfaces, because as I already mentioned, the solution is supposed or was supposed to be used only in protected areas of critical infrastructure, somewhere where you can expect natural surfaces.

So, we will move quickly to the introduction of the second use case, which was focusing on

detection of flooding, to be more specific, for detection of open water flooding in non-urban environments. The solution is using a combination of statistical change detection, pretty similar to the one introduced for the first use case, but with the different settings applied.

And also detection of open water, which is a very specific signature on rather data. There are, again, some several post processing steps applied. One interesting information, mainly, maybe for all of you who are interested in using

greater data for flood detection, our solution is using images from both data polarizations. And if any area is going to be set that it's flooded, it must

be change detected in both polarization and also open water detected in both polarization. So, only the combination of all these four situations lead to a final result of flooding. So, again, to validate the solution, we have chosen five flood events within different European countries in the last few years.

We have processed not only the data for the specific flood events, but for longer time periods, like half a year, to see if there are false positive alarms occurring.

So, it means if the solution is detecting the flooding, although it was no flooding occurring in reality. And, again, to summarize somehow the results, we can say that four of the five flood events were detected successfully.

The only one which was not detected successfully was a flash flood occurring near the city in the Czech Republic. And the developed solution typically provided or is providing flooding pollutants, which are

a bit smaller than those derived from a reference, independent reference data sets. The main reason of this weakness of our product is given by the universal sensitivity parameter settings, which is applied in processing of all locations.

So, it means that you don't have to take the solution and tune the parameters for specific areas for specific events. It's just one universal settings, which can be applied anywhere, anytime.

And it ensures a successful suppression of false alarms. You get a very, very small number of false alarms in any area which we have tested. There is a paper published on this solution, which is describing it in detail and also all

the validation results. So, if this topic interests you, you might be also interested in that publication. And in our limited time, we are forced to move on and to say something about the interferometric solution, which was developed for vertical displacements monitoring over technical infrastructure.

So, we have implemented persistent scatter solution, PS solution. For the PS processing, we are using open source software called SELCIT.

And on the inputs, the user only has to specify the area of interest and the timeline of his interest. And there are also some optional inputs which might be provided if the user wants to do it.

So, the workflow of the algorithm is following after the data selection and its download. There is some standard pre-processing applied and then the own PS processing. And in the post-processing steps, for example, spatial merge of results from individual tracks is being done

or a detection of anomaly points, so of PS points, which behave annually compared to their previous behavior. For the validation, we have realized two sets of works. In the first one, we have established a testing polygon within the campus of our university.

It's consisting of three corner reflectors plus two GNSS stations for an independent evaluation of results.

And the most important thing here is that one of the corner reflectors together with the GNSS antenna, antenna of GNSS receiver, were installed on a construction which allowed us repetitive slow movements.

So, what we were doing was that once per month, we lifted the construction up with the size of 2.5 millimeters.

So, this is that blue line in the right part of the screen where you can see rapid changes of the height once per month. So, in total, after one year of this kind of activity, we have realized a movement of three centimeters

and we were trying to compare with the reality the outputs of the ion star processing from two tracks. Those are these red and black stars in the figure and also with the GNSS data results.

And we can say or we can conclude that the developed solution in this case reached a millimeter level of accuracy of the right estimate. We have more results for it but because of time I'm not able to show to you everything if you're again

interested in this kind of stuff, you can take a look on the paper which is being referenced at this slide. So, I am getting to tell you something about the last validation works.

In this case, we were focusing on displacement monitoring over underground gas storage. We have selected infrastructure in Fredonice within the Czech Republic. We have again installed the GNSS permanent station.

And over the area of gas storage, there is a working cycle where you inject the gas underground during the non-heating season and you take the gas outside, you withdraw it during the heating season. And with this working cycle, there are changes in the terrain height and some movement

can happen which can be detected with techniques as ion star as well as GNSS. So, to show you at least one example of the results, you can see here displacements derived from GNSS data.

These are these orange points and from the ion star processing, these are these blue points. And you can clearly see here a periodic behavior with a period of approximately one year

when the terrain is not moving only up and down with the amplitude of approximately one centimeter. But in the east-west direction, the east-west component of the displacement is shown in the bottom part of the figure.

Okay, to conclude, within the ESA funded project, we have developed automatic processing change for Sentinel-1 imagery, for use cases, all developed solutions were validated against independent reference datasets.

Two of the processing chains, one for the flood detection and the PES InSAR solution were implemented in the SATSIDE platform which was developed within the project by CGI-IT company.

And there will be a next presentation given by Wojciech Run from this company who will tell more about this, about their works. So, you are welcome to see it. And that's all from my side. I want to thank you for your attention. If there are any questions, I would be happy to answer them.