Latest / Satellite image deep learning / PhiDown: Fast, Simple Access to Copernicus Data

Transcript

PhiDown: Fast, Simple Access to Copernicus Data · Satellite image deep learning

• 0:04
  SPEAKER_01: Hi, Roberto.
• 0:05
  SPEAKER_01: Welcome back to the podcast.
• 0:07
  SPEAKER_00: Great.
• 0:08
  SPEAKER_00: It's a pleasure to be back on this podcast.
• 0:11
  SPEAKER_01: Fantastic.
• 0:11
  SPEAKER_01: Why don't you tell us a little bit about yourself and where you work?
• 0:14
  SPEAKER_00: So I'm an internal research fellow.
• 0:19
  SPEAKER_00: It's like a postdoc position.
• 0:22
  SPEAKER_00: So I'm a researcher and I work in the Philabs in ESRIN,
• 0:26
  SPEAKER_00: which is the research center in Frascati of the European Space Agency.
• 0:30
  SPEAKER_01: Brilliant.
• 0:31
  SPEAKER_01: And you've been involved in a number of different projects,
• 0:33
  SPEAKER_01: but today we're going to talk about FIDAM.
• 0:35
  SPEAKER_01: Why don't you give us the intro to that?
• 0:38
  SPEAKER_00: FIDAM is mostly an answer, okay?
• 0:42
  An answer to many people that wanted to access lots and lots of
• 0:46
  data at different
• 0:48
  levels and had a little bit of problem interfacing with the
• 0:52
  Copernicus Data Space
• 0:54
  SPEAKER_00: Ecosystem API.
• 0:55
  SPEAKER_00: Right.
• 0:56
  SPEAKER_00: So many people come at me and to my colleagues
• 1:00
  SPEAKER_00: asking for a solution to this problem.
• 1:02
  SPEAKER_00: And I will always reminding them that there is this API that existing,
• 1:06
  SPEAKER_00: that were already existing,
• 1:08
  SPEAKER_00: but they were a bit complex to use.
• 1:10
  So CDOWN is simply a wrapper around that API that allows people
• 1:17
  to use the OData
• 1:18
  SPEAKER_00: API,
• 1:19
  SPEAKER_00: but without the struggle of learning how to use the OData API.
• 1:23
  SPEAKER_01: Right.
• 1:24
  SPEAKER_01: So the struggles were authentication or something like that?
• 1:29
  SPEAKER_00: It's mostly related to the documentation and go through it,
• 1:33
  SPEAKER_00: understand how does it work,
• 1:35
  SPEAKER_00: the OData API scheme,
• 1:37
  SPEAKER_00: how to set the parameters correctly,
• 1:39
  SPEAKER_00: and stuff like that.
• 1:42
  SPEAKER_01: Right.
• 1:42
  So FIDOWN is a tool written to interface with their API,
• 1:47
  which greatly simplifies the use of it.
• 1:50
  SPEAKER_00: Yeah, exactly.
• 1:52
  SPEAKER_00: And most importantly,
• 1:54
  SPEAKER_00: it's built with the S5 protocol,
• 1:57
  which is new and allows people to download the data in a very
• 2:01
  fast and efficient
• 2:01
  SPEAKER_00: way.
• 2:02
  SPEAKER_00: It's very made up to be simple and to scale up your productivity.
• 2:07
  SPEAKER_01: Right.
• 2:08
  SPEAKER_01: I'd love to learn a bit more about that.
• 2:09
  SPEAKER_01: It sounds a bit like S3, but it's obviously something different.
• 2:12
  SPEAKER_00: Yeah, it is derived from S3, actually.
• 2:16
  It's the new protocol,
• 2:17
  which is like a bit five,
• 2:19
  ten times more faster than S3.
• 2:22
  SPEAKER_01: Right.
• 2:23
  SPEAKER_01: Okay.
• 2:24
  And so is that a backend implementation or is that something
• 2:29
  that impacts people using the tool?
• 2:33
  It's just something that's empowering people to use the S5
• 2:37
  through the Copernicus
• 2:39
  SPEAKER_00: Data Space ecosystem.
• 2:41
  SPEAKER_01: Okay.
• 2:42
  SPEAKER_01: And digging into the functionality of the tool,
• 2:45
  SPEAKER_01: how does it compare to other tools that I already know about,
• 2:47
  SPEAKER_01: like Google Earth Engine or Stack?
• 2:50
  SPEAKER_00: And that's a good point.
• 2:52
  This is a question that is mostly raised when we speak about
• 2:55
  Fidam because there
• 2:57
  SPEAKER_00: are a lot of tools and they are very valuable.
• 2:59
  SPEAKER_00: They do a lot of things.
• 3:00
  SPEAKER_00: And I think that they serve probably 90% of the users.
• 3:05
  SPEAKER_00: But these tools also, they come with some limitations.
• 3:08
  SPEAKER_00: For instance,
• 3:09
  if you would like to download a certain kind of level of a
• 3:12
  specific product,
• 3:14
  SPEAKER_00: you cannot simply do through Google Earth Engine.
• 3:17
  SPEAKER_00: So for instance,
• 3:17
  SPEAKER_00: if you like to access to SLC data,
• 3:20
  SPEAKER_00: or if you want to go even more beating down,
• 3:23
  SPEAKER_00: if you need to download level zero data,
• 3:25
  SPEAKER_00: or if you need to do...
• 3:27
  SPEAKER_00: different kind of analysis to test specific algorithms.
• 3:32
  In that case,
• 3:32
  you need to download a large amount of data,
• 3:35
  of low level data.
• 3:37
  SPEAKER_00: It's not simply that usually you can do on Google Earth Engine.
• 3:40
  SPEAKER_01: Right.
• 3:41
  So the Google Earth Engine obviously doesn't index all the
• 3:43
  kinds of levels.
• 3:45
  SPEAKER_01: You've mentioned that Copernicus does index all of them.
• 3:48
  SPEAKER_01: What about the format of the data that you get?
• 3:50
  SPEAKER_01: Is it familiar structures?
• 3:53
  There is a difference between this and Google Earth Engine and
• 3:56
  other tools because
• 3:57
  SPEAKER_00: they use the stack format.
• 3:58
  So basically you have like a sort of mosaic and you can access
• 4:02
  different portion of
• 4:03
  SPEAKER_00: the earth and simply looking at that.
• 4:05
  Yeah.
• 4:05
  Well,
• 4:05
  feed down access to the products as they are stored on the
• 4:10
  cloud server.
• 4:11
  It means that each product can be downloaded as it is in its
• 4:16
  proprietary format of ESA.
• 4:19
  SPEAKER_00: So you need to process it using the right tools.
• 4:23
  SPEAKER_00: It's not ready plug and play.
• 4:25
  SPEAKER_00: But it also comes with a little,
• 4:27
  SPEAKER_00: a bit more of metadata and additional information that usually are not
• 4:32
  SPEAKER_00: self-contained in the JOT that you can download through other tools.
• 4:36
  SPEAKER_00: So you can see really a lot of power to do certain things.
• 4:41
  SPEAKER_00: But of course, you need to know how to manage properly that products.
• 4:45
  SPEAKER_01: Right.
• 4:45
  SPEAKER_01: Okay.
• 4:46
  SPEAKER_01: So it will synchronize the data that you want to your local location,
• 4:50
  but then it's your responsibility to know how to further
• 4:54
  process it and access it.
• 4:57
  SPEAKER_00: Exactly.
• 4:57
  SPEAKER_00: Exactly.
• 4:58
  SPEAKER_01: Brilliant.
• 4:59
  SPEAKER_01: And what kind of uses have people got for FIDOWN already?
• 5:05
  SPEAKER_00: So scaling up your productivity, for instance.
• 5:08
  So if you need a chain that autonomously downloaded the newest
• 5:13
  product gathered by
• 5:15
  SPEAKER_00: one of the Sentinel missions and autonomously put it in your workflow,
• 5:19
  SPEAKER_00: you can do that with feed-on or accessing to low-level products.
• 5:25
  SPEAKER_00: which are necessary for doing certain kind of analysis.
• 5:30
  SPEAKER_00: Other things that you can do is creating data sets,
• 5:33
  which are very important for the modern world with these
• 5:36
  data-hungry algorithms
• 5:38
  SPEAKER_00: like foundation models.
• 5:40
  SPEAKER_00: And we know how many data we need nowadays.
• 5:44
  It's very easy to download the data in similar locations coming
• 5:50
  and spanning from
• 5:51
  SPEAKER_00: different sensors.
• 5:54
  I know that in your own research,
• 5:55
  you've used the level zero for timeliness.
• 5:58
  SPEAKER_01: Is that right?
• 6:00
  SPEAKER_00: We use level zero for different applications.
• 6:03
  I think there's still a lot of information that we can gather
• 6:07
  directly from the level zero data.
• 6:09
  SPEAKER_00: And it's still a topic which is underexplored.
• 6:12
  SPEAKER_00: But if you think about it,
• 6:13
  SPEAKER_00: the value that we have is not in the product per se,
• 6:16
  SPEAKER_00: but it's in information that we have in the products,
• 6:19
  SPEAKER_00: right?
• 6:20
  So why don't we train deep learning models to extract that
• 6:24
  information directly at level zero?
• 6:26
  SPEAKER_01: Right.
• 6:28
  SPEAKER_01: And that's exactly what you've done, right?
• 6:31
  SPEAKER_00: We've done this for several projects,
• 6:34
  SPEAKER_00: one employing like wildfire detection,
• 6:37
  SPEAKER_00: another for vessels.
• 6:39
  SPEAKER_00: We proved this for several missions, including Sentinel-2 and Venus.
• 6:44
  SPEAKER_01: Right.
• 6:45
  So Fire Down would fit into the workflow that would enable you
• 6:48
  to access those
• 6:49
  SPEAKER_01: products as fast as feasibly possible.
• 6:53
  SPEAKER_00: Yeah, that's the idea.
• 6:55
  SPEAKER_01: Brilliant.
• 6:56
  SPEAKER_01: I'm curious, this is quite a recent development.
• 6:59
  I'm just wondering how you developed it in light of what
• 7:03
  everyone's doing with LLMs,
• 7:04
  et cetera.
• 7:08
  SPEAKER_00: how I developed it.
• 7:09
  SPEAKER_00: That's a good point.
• 7:11
  SPEAKER_00: Now the workflow is changing a little bit in last year.
• 7:13
  As you may notice that there are a lot of LLMs and the
• 7:17
  autocoding are spreading.
• 7:20
  SPEAKER_00: I think that nowadays everyone is using these tools,
• 7:24
  but it's very important that someone uses this tool with having
• 7:28
  knowledge of what
• 7:28
  he's doing because simply assigning a task to a coding agent to
• 7:33
  unsolve your
• 7:34
  SPEAKER_00: problem.
• 7:35
  SPEAKER_01: No.
• 7:35
  But we can discuss about this,
• 7:38
  but I think that nowadays everyone is using coding agents.
• 7:42
  SPEAKER_01: Yeah.
• 7:42
  It feels like it would be a good use case for it to sort of
• 7:45
  create this tool to
• 7:46
  SPEAKER_01: access an API that is already well documented.
• 7:49
  SPEAKER_00: Yeah.
• 7:49
  I think that there is someone that asked me to create an MCP to
• 7:53
  accessing the
• 7:56
  SPEAKER_00: Copernicus Autospace Ecosystem Resources.
• 7:58
  SPEAKER_00: And maybe I'm going to do that in the future.
• 8:01
  SPEAKER_01: Exciting.
• 8:01
  SPEAKER_01: Right.
• 8:01
  SPEAKER_01: So this Firedown is on GitHub.
• 8:03
  SPEAKER_01: I'm just having a look at this review now.
• 8:04
  SPEAKER_01: It's very nicely documented.
• 8:07
  SPEAKER_01: You've also got an example notebook to get started.
• 8:10
  So what we're going to do is jump to a separate recording and
• 8:14
  we're going to run
• 8:15
  SPEAKER_01: through that with yourself.
• 8:18
  SPEAKER_01: That will be on YouTube.
• 8:20
  SPEAKER_01: Just as we wrap things up here,
• 8:21
  is there anything else people should know about Firedown or if
• 8:24
  people want to get
• 8:25
  SPEAKER_01: involved with it?
• 8:26
  SPEAKER_01: What's your advice?
• 8:29
  My advice is to just try it and leave me a message or open up a
• 8:34
  request if there is
• 8:35
  SPEAKER_00: a feature that we really like.
• 8:37
  SPEAKER_00: I tried to put everything on our documentation.
• 8:39
  So even the attributes that you need to put for searching
• 8:43
  specific products.
• 8:44
  But I cannot cover all the use case because I cannot work in
• 8:48
  all the applications.
• 8:49
  So really,
• 8:49
  this project is something that is born from the community,
• 8:53
  from a community demand.
• 8:54
  SPEAKER_00: And it's done for the community.
• 8:56
  SPEAKER_00: Okay.
• 8:57
  So it's really something that I would like to build together
• 9:00
  with all the community.
• 9:02
  SPEAKER_01: Brilliant.
• 9:02
  SPEAKER_01: Well, I wish you every success of that.
• 9:05
  SPEAKER_01: All right.
• 9:05
  SPEAKER_01: With that, we'll wrap things up and move to the next video.
• 9:08
  SPEAKER_01: So thanks.
• 9:09
  SPEAKER_01: Thanks again, Roberto.
• 9:11
  SPEAKER_00: Thank you, Robin.