Latest / Satellite image deep learning / Challenges and opportunities for Ai mapping

Transcript

Challenges and opportunities for Ai mapping · Satellite image deep learning

• 0:05
  SPEAKER_01: Hi, Toby.
• 0:06
  SPEAKER_01: Welcome to the podcast.
• 0:08
  SPEAKER_00: Yeah.
• 0:08
  SPEAKER_00: Hi, Robert.
• 0:09
  SPEAKER_00: Thank you for having me.
• 0:11
  SPEAKER_01: It's my pleasure.
• 0:12
  SPEAKER_01: All right.
• 0:12
  SPEAKER_01: Let's begin with where you work and what you do.
• 0:16
  SPEAKER_00: So, yeah, my name is Tobias Augsburger.
• 0:19
  SPEAKER_00: I'm working for Open Energy Transition.
• 0:22
  SPEAKER_00: I'm leading the grid mapping activity.
• 0:26
  um we just started a small initiative called um map your grid
• 0:32
  uh that is funded
• 0:33
  actually by break for energy um that's the company of bill
• 0:37
  gates that tries to do a
• 0:39
  SPEAKER_00: lot in the energy space um
• 0:44
  So Bill Gates actually envisioned to create an open source
• 0:48
  operating system in the
• 0:49
  SPEAKER_00: energy sector.
• 0:51
  And we are supporting this with different open source
• 0:54
  developments within this company.
• 0:57
  um yeah i have a background in actually aerospace engineering
• 1:02
  so actually i'm an
• 1:03
  aerospace engineer i worked in aerodynamics for quite some time
• 1:07
  instrumentation
• 1:08
  remote sensing so remote sensing instrumentation then i
• 1:12
  switched to um evv codes
• 1:16
  and automation of evv codes did a lot of with open source and
• 1:19
  robotics so i really
• 1:20
  SPEAKER_00: got into open source actually by
• 1:23
  SPEAKER_00: using it in robotics,
• 1:25
  automating big trucks to move containers on a yard was
• 1:32
  something we did for DHL for
• 1:34
  SPEAKER_00: quite some time with open source.
• 1:37
  And then I lost my job and did the open sustainable technology
• 1:43
  project as a COVID project.
• 1:47
  SPEAKER_00: Yeah, right.
• 1:48
  SPEAKER_01: So you had a bit of time.
• 1:49
  SPEAKER_01: You thought, I'll get into open source.
• 1:52
  SPEAKER_01: Interesting.
• 1:53
  And so there's kind of an arc here towards understanding energy
• 1:56
  infrastructure and
• 1:59
  SPEAKER_01: even sort of making sense of that on a bigger scale.
• 2:03
  SPEAKER_00: Exactly.
• 2:03
  SPEAKER_00: Yes.
• 2:05
  We really try to create more or less a digital twin of the
• 2:09
  global electrical grid.
• 2:11
  SPEAKER_00: Yeah.
• 2:12
  SPEAKER_00: All right.
• 2:12
  SPEAKER_01: Okay.
• 2:12
  SPEAKER_01: Yeah.
• 2:14
  SPEAKER_01: Fascinating.
• 2:15
  SPEAKER_01: So just what are we talking about?
• 2:17
  We're talking about pylons,
• 2:19
  we're talking about substations,
• 2:21
  that kind of infrastructure,
• 2:22
  right?
• 2:24
  exactly um so we map the lines because you can also see the
• 2:29
  lines and satellite
• 2:30
  imagery so our main source uh for mapping or raw data for
• 2:36
  mapping is satellite
• 2:37
  imagery and we can see the pylons or power towers we can see
• 2:42
  substations and we can
• 2:44
  SPEAKER_00: see even the transmission lines in many cases
• 2:47
  right right so literally those wires across the landscape that
• 2:50
  are probably thin as
• 2:52
  what your wrist or something um exactly yeah we really uh we
• 3:00
  normally start with
• 3:01
  the substations and then from that you see all the wires going
• 3:05
  through all the
• 3:06
  SPEAKER_00: regions
• 3:09
  SPEAKER_00: connecting actually the energy systems across all the countries.
• 3:16
  SPEAKER_01: Right,
• 3:18
  SPEAKER_01: so high resolution imagery,
• 3:20
  SPEAKER_01: you look for the substations,
• 3:21
  SPEAKER_01: you know that these things are connected by the wires.
• 3:24
  SPEAKER_01: Sounds pretty straightforward.
• 3:26
  I think there's plenty of machine learning models out there
• 3:28
  that can segment objects.
• 3:30
  SPEAKER_01: What are the challenges there?
• 3:33
  SPEAKER_00: Yeah,
• 3:33
  SPEAKER_00: we,
• 3:34
  SPEAKER_00: we use,
• 3:36
  SPEAKER_00: um,
• 3:36
  SPEAKER_00: mainly at the moment,
• 3:38
  SPEAKER_00: um,
• 3:39
  SPEAKER_00: human beings actually that are,
• 3:41
  SPEAKER_00: uh,
• 3:42
  SPEAKER_00: uh,
• 3:42
  SPEAKER_00: mapping the grid and open street maps.
• 3:44
  SPEAKER_00: So 70% or roughly about 70% of the grid has,
• 3:48
  SPEAKER_00: uh,
• 3:48
  SPEAKER_00: mostly actually been mapped over the last 17 years by,
• 3:53
  SPEAKER_00: um,
• 3:53
  SPEAKER_00: human beings.
• 3:54
  SPEAKER_00: Um,
• 3:56
  SPEAKER_00: But we are now,
• 3:58
  SPEAKER_00: let's say on the last 30%,
• 3:59
  we are now also investigating machine learning approaches to
• 4:07
  detect substations.
• 4:08
  That's the most easiest object you can,
• 4:12
  and biggest object also you can find in the grid.
• 4:19
  SPEAKER_00: But we also investigating to detect the next power tower
• 4:25
  SPEAKER_00: But this comes with a lot of challenges.
• 4:27
  SPEAKER_01: Right,
• 4:30
  SPEAKER_01: so substations,
• 4:31
  SPEAKER_01: quite large,
• 4:32
  SPEAKER_01: regular shapes,
• 4:33
  SPEAKER_01: you think those are pretty straightforward,
• 4:35
  SPEAKER_01: but you have other infrastructure,
• 4:36
  SPEAKER_01: which is more challenging.
• 4:38
  And that's why the human oversight is required to kind of
• 4:41
  connect the dots and
• 4:43
  SPEAKER_01: identify all these things as being part of the same network.
• 4:46
  SPEAKER_01: Is that right?
• 4:47
  SPEAKER_00: Yeah, that's right.
• 4:48
  But it's also the case that you find a lot of different grid
• 4:54
  infrastructure that
• 4:55
  SPEAKER_00: has actually been synthetically created by AI.
• 4:59
  And at the end,
• 5:02
  the challenge is to harmonize all this data into one map that
• 5:05
  is validated.
• 5:07
  SPEAKER_00: And this is where the human being comes into the process.
• 5:13
  SPEAKER_00: Interesting.
• 5:15
  Yeah,
• 5:16
  you have for one country sometimes five or six different grid
• 5:19
  maps or grid data sets.
• 5:23
  SPEAKER_00: And what we normally do, we don't copy paste anything in OpenStreetMap.
• 5:28
  SPEAKER_00: That's very important to understand.
• 5:30
  SPEAKER_00: We use them as a hint layer.
• 5:33
  so we have our open street map editors and then we show the
• 5:39
  open street map data
• 5:40
  and then we put on another layer actually this ai generated or
• 5:45
  however this data
• 5:46
  has been generated and then we have to integrate this data set
• 5:52
  by validating every
• 5:53
  data point right so power tower got it so there's basically
• 5:59
  false positives from
• 6:00
  SPEAKER_01: the machine learning models is that what you're saying
• 6:04
  um yes it's about false positives um but it's really also about
• 6:10
  the connections so
• 6:12
  um for example it is relevant um at which side of the
• 6:18
  substation a transmission
• 6:19
  SPEAKER_00: line is connected to estimate the voltage of the line it's sometimes very
• 6:27
  SPEAKER_00: interesting to um also see what is
• 6:32
  um on the end of a substation and this is also hard to uh to do
• 6:36
  with AI because you
• 6:38
  find sometimes very different infrastructure at the end you
• 6:42
  find bigger loads you
• 6:43
  find sometimes just a distribution substation of the city you
• 6:48
  find sometimes a
• 6:49
  SPEAKER_00: generate a big car plant and
• 6:54
  this final mapping process of the grid allows us to bring in
• 7:01
  all this context into
• 7:03
  SPEAKER_00: the data set.
• 7:04
  SPEAKER_00: So it's really a journey that you have in grid mapping.
• 7:11
  SPEAKER_00: And with machine learning, we would not have this.
• 7:14
  SPEAKER_00: So we could not bring all this knowledge into the grid.
• 7:19
  SPEAKER_00: We would not have the opportunity also to
• 7:26
  bring metadata into the grid that demands a certain context
• 7:33
  understanding of the grid.
• 7:35
  So there's a lot of different layers and context understanding
• 7:39
  that you need to map the grid.
• 7:44
  SPEAKER_01: I guess my next question is,
• 7:46
  presumably you could train models to do things like identify
• 7:48
  which side of wire was
• 7:50
  connected and start to build up some of this context itself or
• 7:53
  input those features
• 7:55
  to another model that could make predictions about how these
• 7:59
  things are connected
• 8:00
  SPEAKER_01: or which class they are.
• 8:04
  SPEAKER_00: Let's say this way.
• 8:05
  SPEAKER_00: It's technically possible, yes, but there are limitations to that.
• 8:13
  one is just the data availability for machine learning of this
• 8:16
  high resolution data.
• 8:19
  SPEAKER_00: So in OpenStreetMap,
• 8:22
  SPEAKER_00: we have S3 world imagery,
• 8:25
  SPEAKER_00: we have Bing aerial imagery,
• 8:28
  but this data is just being provided for the OpenStreetMap
• 8:32
  community for free to do
• 8:34
  SPEAKER_00: manual work mapping.
• 8:36
  SPEAKER_01: Right.
• 8:39
  SPEAKER_00: Every machine learning approach,
• 8:41
  we would actually need to buy the satellite imagery and not
• 8:47
  just one imagery
• 8:49
  SPEAKER_00: product.
• 8:50
  SPEAKER_00: And that's the challenge.
• 8:51
  SPEAKER_00: So for us as a human to validate this data,
• 8:54
  SPEAKER_00: you have to switch between sometimes four different satellite imagery.
• 8:59
  SPEAKER_01: Right.
• 9:01
  Because the transmission lines in,
• 9:04
  for example,
• 9:05
  in Africa,
• 9:06
  they go through areas where
• 9:09
  age of the satellite imagery and the quality of the satellite
• 9:12
  imagery is changing a
• 9:13
  lot so you go 100 kilometers then you have to change to another
• 9:18
  satellite imagery
• 9:19
  product then you have to change again you have change again and
• 9:22
  sometimes for us we
• 9:23
  really have to switch four times between the same imagery to
• 9:28
  ensure that this is
• 9:29
  SPEAKER_00: actually a power tower right
• 9:33
  So you would basically have four times the problem for a model
• 9:36
  because typically a
• 9:36
  SPEAKER_01: model is trained on one kind of imagery,
• 9:39
  SPEAKER_01: one kind of target.
• 9:40
  SPEAKER_01: But here we have a lot more diversity to handle.
• 9:43
  SPEAKER_01: Still sounds like maybe a solvable problem for machine learning,
• 9:47
  SPEAKER_01: but I'm getting a good impression of the sort of key challenges.
• 9:52
  SPEAKER_00: It definitely is.
• 9:53
  There is a very interesting open source prototype that has been
• 9:58
  developed by a
• 10:00
  SPEAKER_00: group in New York from Lindsay Labs.
• 10:05
  SPEAKER_00: And they do substation detection using Sentinel-2.
• 10:11
  And since Sentinel-2 is an open data set,
• 10:15
  it's allowed as a raw data source for OpenStreetMap.
• 10:22
  And we are investigating this because one of the main
• 10:28
  challenges is to have the substation.
• 10:30
  If you have the substations of a country,
• 10:33
  I would say this is 50% of the work.
• 10:37
  SPEAKER_00: Then you just have to connect the dots.
• 10:42
  SPEAKER_00: And this is possible using...
• 10:49
  all the different spectres of the uh of centenary two right
• 10:54
  right so there was a
• 10:55
  solution already developed i think i've seen that a little
• 10:57
  while back um how
• 10:59
  SPEAKER_00: effective was it i think it the false positives were very little um the
• 11:07
  segmentation was uh the intersection of union was not perfect
• 11:11
  but still very good
• 11:13
  SPEAKER_00: at least from the paper and from the um
• 11:18
  uh from the code i've seen uh it works actually very well um
• 11:23
  and it would be for us
• 11:25
  good enough because even if you would say even if you would
• 11:28
  have i don't know uh 10
• 11:31
  false positives since we are this has been validated by a human
• 11:36
  being at the end
• 11:37
  SPEAKER_00: it's not a problem
• 11:40
  But it's really key for us because every power plant comes with
• 11:44
  a substation.
• 11:46
  SPEAKER_00: Every bigger industrial loads comes with a substation.
• 11:52
  You have actually the whole power infrastructure if you have a
• 11:55
  good substation detection.
• 11:56
  SPEAKER_01: All right.
• 11:57
  So that sounds like a pretty promising model,
• 12:00
  just for that particular task.
• 12:03
  SPEAKER_01: Does it work globally?
• 12:04
  Are there some reasons why it can't be deployed for this
• 12:09
  initial vetting of where
• 12:12
  SPEAKER_01: the substations are?
• 12:13
  SPEAKER_00: So the training data set...
• 12:19
  that has been created for this is using OpenStreetMap data as a
• 12:25
  ground truth.
• 12:25
  SPEAKER_00: So OpenStreetMap is used a lot as ground truth validated data set.
• 12:30
  And I used actually the polygons and projected then on Sentinel
• 12:35
  two data and use
• 12:37
  SPEAKER_00: this,
• 12:37
  SPEAKER_00: I think 50,000 substations all around the world.
• 12:41
  SPEAKER_00: Yeah, the results were pretty good.
• 12:50
  and i think the only limitation that you actually have is areas
• 12:57
  where there's just
• 12:58
  a lot of clouds it's always right so optical data set some
• 13:02
  places amazon or
• 13:04
  wherever lots of clouds potentially you need more than one
• 13:09
  modality to detect the
• 13:10
  substations globally yes i we investigated also sentinel one
• 13:16
  data if it might be
• 13:18
  SPEAKER_00: possible
• 13:23
  SPEAKER_00: I think it could be done.
• 13:28
  And it's,
• 13:29
  it,
• 13:29
  it,
• 13:30
  but it also depends on the size of the substation itself.
• 13:34
  SPEAKER_00: Like this really high voltage substations.
• 13:37
  SPEAKER_00: They are so big.
• 13:39
  Um,
• 13:39
  you can actually work with very low resolution satellite data
• 13:43
  to try and detect them.
• 13:45
  SPEAKER_00: Um, but those are actually easy to find for this.
• 13:50
  SPEAKER_00: You don't have detection.
• 13:52
  SPEAKER_00: Um,
• 13:55
  The harder ones are the small distribution substations that are
• 13:59
  at the end of the grid.
• 14:03
  SPEAKER_00: But those are especially interested because the interesting for us,
• 14:10
  because the distribution grid is something that you cannot map
• 14:18
  very well with even
• 14:21
  SPEAKER_00: with high resolution satellite data.
• 14:25
  SPEAKER_00: because it's mostly in urban areas,
• 14:28
  SPEAKER_00: you have the poles sometimes,
• 14:30
  SPEAKER_00: you can't see them,
• 14:31
  SPEAKER_00: but normally you see poles based on the shadow or power towers,
• 14:35
  you see always based on the shadow of those infrastructure
• 14:38
  because they are so
• 14:39
  SPEAKER_00: tall.
• 14:40
  SPEAKER_00: And in this urban areas, it's sometimes pretty hard to detect this.
• 14:46
  SPEAKER_00: But if you have the distribution substations,
• 14:49
  and it can use again machine learning and try and this is
• 14:53
  there's a lot of a lot of
• 14:55
  groups trying to estimate the distribution grid using machine
• 15:00
  learning approaches
• 15:01
  SPEAKER_00: and this can be um supported a lot if you know the nodes the distribution
• 15:09
  SPEAKER_00: substation that are the nodes of the distribution grid
• 15:13
  SPEAKER_01: Right.
• 15:14
  So this sort of comes back to your earlier point that the
• 15:16
  experts use multiple
• 15:17
  SPEAKER_01: sources of imagery.
• 15:18
  SPEAKER_01: They've done a lot of personal training.
• 15:20
  SPEAKER_01: They have an understanding of how grid networks work.
• 15:23
  And they are able to identify these smaller substations in more
• 15:27
  challenging areas.
• 15:28
  SPEAKER_01: But currently, machine learning methods are not there.
• 15:34
  SPEAKER_00: We have sort of the substation detection that has been developed so far.
• 15:43
  SPEAKER_00: using mostly bigger substations, yes.
• 15:46
  SPEAKER_01: Yeah.
• 15:46
  SPEAKER_01: Okay, brilliant.
• 15:48
  So I think that really helps sort of clarify where the state of
• 15:50
  the art is and what
• 15:51
  SPEAKER_01: the sort of frontier looks like and what the challenges are.
• 15:55
  SPEAKER_01: Obviously, you've been working to curate a lot of truth data.
• 15:59
  SPEAKER_01: As you mentioned,
• 16:00
  SPEAKER_01: OpenStreetMap is often used for generating,
• 16:04
  SPEAKER_01: you know,
• 16:04
  SPEAKER_01: annotated data sets for training models.
• 16:06
  SPEAKER_01: Yes.
• 16:08
  I'm not sure how many people were aware of that and also the
• 16:10
  richness of the data
• 16:11
  SPEAKER_01: that you've curated.
• 16:13
  So I think one thing we want to do is just highlight that that
• 16:15
  resource exists and
• 16:17
  SPEAKER_01: is being developed for this purpose,
• 16:19
  SPEAKER_01: right?
• 16:21
  SPEAKER_00: Yes,
• 16:23
  SPEAKER_00: it's not...
• 16:24
  SPEAKER_00: It's used for machine learning a lot,
• 16:28
  SPEAKER_00: but it's also used for energy system modeling,
• 16:30
  SPEAKER_00: like actually half a ground truth data set.
• 16:34
  SPEAKER_00: It's used a lot...
• 16:38
  SPEAKER_00: for regional studies, for larger studies.
• 16:40
  SPEAKER_00: It's used by the World Bank to actually,
• 16:45
  SPEAKER_00: to improve the electrification of,
• 16:51
  SPEAKER_00: for example,
• 16:52
  SPEAKER_00: areas in Africa and South America.
• 16:56
  SPEAKER_00: So this data is really one
• 16:59
  actually the only data set that is standardized for the
• 17:03
  electrical grid worldwide.
• 17:05
  Many low to middle income countries,
• 17:06
  there's actually no good data about the electrical grid.
• 17:12
  And this is a real problem for investors and for the expansion
• 17:18
  of this electrical grid.
• 17:22
  And yeah,
• 17:23
  OpenStreetMap is used as one of the best sources because it's
• 17:28
  validated.
• 17:29
  SPEAKER_00: And it's accepted by the community,
• 17:33
  because it's used for years by larger organizations like the
• 17:37
  World Bank to do
• 17:38
  SPEAKER_00: investments into the grid.
• 17:39
  SPEAKER_01: Right.
• 17:40
  SPEAKER_01: So significant uses for this data set,
• 17:45
  SPEAKER_01: you know,
• 17:45
  SPEAKER_01: not just for machine learning and expanding the data set,
• 17:48
  SPEAKER_01: but also for understanding infrastructure in a country,
• 17:51
  SPEAKER_01: enabling investment,
• 17:54
  SPEAKER_01: etc.
• 17:55
  SPEAKER_00: Exactly, yeah.
• 17:56
  SPEAKER_01: Brilliant.
• 17:58
  SPEAKER_01: And interesting constraints,
• 17:59
  SPEAKER_01: as you mentioned,
• 18:00
  SPEAKER_01: around sort of licensing of data for particular uses.
• 18:03
  And then also when you try and deploy these generated locations
• 18:09
  into OpenStreetMap,
• 18:10
  SPEAKER_01: there's certain criteria that must meet,
• 18:13
  like you must be able to independently validate those results
• 18:17
  using open data.
• 18:17
  So it's a really interesting context there that kind of guides
• 18:21
  the solutions people
• 18:22
  SPEAKER_01: might be looking to develop.
• 18:25
  SPEAKER_01: Brilliant.
• 18:25
  SPEAKER_01: That's a really interesting conversation.
• 18:27
  SPEAKER_01: I think it's a very interesting problem as well.
• 18:28
  SPEAKER_01: As you mentioned,
• 18:29
  SPEAKER_01: it's kind of not just detect a thing,
• 18:31
  SPEAKER_01: but understand how it fits into a bigger network and challenges like the
• 18:36
  SPEAKER_01: multimodality potentially helping.
• 18:38
  And I think lots of potential there for machine learning and
• 18:43
  really highlights kind
• 18:44
  SPEAKER_01: of the limitations of the current technology.
• 18:47
  SPEAKER_01: I guess.
• 18:48
  SPEAKER_01: In terms of just to sort of bring things to a head,
• 18:51
  where are the main platforms where this information is
• 18:54
  communicated and where you
• 18:56
  SPEAKER_01: share your own kind of resources on this topic?
• 19:01
  so um i think one of the best um websites to actually have a
• 19:07
  look on this data is
• 19:08
  um open infrastructure map or you find it at openinframap.org
• 19:15
  and this is a
• 19:17
  visualization of the global electrical grid data this is an
• 19:21
  open street map
• 19:22
  SPEAKER_00: including power plants substations and um
• 19:28
  SPEAKER_00: but also all the underground infrastructure.
• 19:33
  SPEAKER_00: So you will find,
• 19:36
  SPEAKER_00: especially for Europe and the US,
• 19:38
  SPEAKER_00: this data is from data quality,
• 19:44
  SPEAKER_00: similar to official data from system operators.
• 19:51
  If you visit,
• 19:51
  if you want to improve this data,
• 19:53
  you would visit map your grid.org.
• 19:57
  SPEAKER_00: That's our website.
• 19:59
  SPEAKER_00: Um,
• 20:00
  SPEAKER_00: here you can find our starter kit or videos,
• 20:03
  but also some tooling that is just simplifying and automating
• 20:07
  the process around
• 20:09
  SPEAKER_00: transmission grid mapping.
• 20:11
  SPEAKER_00: So we, we try to standardize a lot around this topic.
• 20:16
  SPEAKER_00: Um,
• 20:18
  But also the OpenStreetMap Wikipedia page has a lot of very
• 20:23
  interesting content.
• 20:28
  You will find it actually if you search for OpenStreetMap Wiki
• 20:32
  and Power Networks.
• 20:36
  Debbie will find like a landing page that will give you an
• 20:40
  introduction into also
• 20:42
  SPEAKER_00: the topic,
• 20:43
  but also the local pages from the local groups of people that
• 20:52
  are focusing on
• 20:54
  SPEAKER_00: curating,
• 20:55
  SPEAKER_00: improving the grid data for their country.
• 20:58
  SPEAKER_00: And they are collecting data there on their country pages.
• 21:02
  SPEAKER_00: So I think this is the main
• 21:05
  SPEAKER_00: um, websites, um, when it comes to open street map.
• 21:10
  It's great to hear about these,
• 21:13
  these global communities working to map their areas.
• 21:16
  And I just think it's an amazing resource that the community
• 21:20
  has put together.
• 21:22
  Yes,
• 21:23
  it's really,
• 21:24
  uh,
• 21:25
  I must personally say I'm really,
• 21:27
  um,
• 21:28
  uh,
• 21:30
  impressed by the data quality.
• 21:31
  SPEAKER_00: Um,
• 21:34
  It's a little bit hard to prove the data quality within
• 21:37
  OpenSuiteMap and we are
• 21:38
  SPEAKER_00: investigating this at the moment.
• 21:40
  SPEAKER_00: Also using machine learning,
• 21:41
  SPEAKER_00: do random sample approach to take data,
• 21:44
  SPEAKER_00: try if you can detect it like it's been mapped.
• 21:47
  SPEAKER_00: Mm-hmm.
• 21:49
  SPEAKER_00: But it's really,
• 21:50
  SPEAKER_00: yeah,
• 21:52
  what I think this is something very special that a lot of
• 21:56
  people don't really
• 21:57
  SPEAKER_00: understand.
• 21:58
  SPEAKER_00: How do you curate a data set over 17 years?
• 22:03
  SPEAKER_00: What is the strategy to improve this data on and on and on,
• 22:08
  SPEAKER_00: especially when it comes to machine learning,
• 22:10
  SPEAKER_00: where there's a lot of people that actually are generating data that
• 22:18
  is being potentially a source,
• 22:22
  but someone at the end still needs to harmonize all of this.
• 22:29
  SPEAKER_00: And OpenStreetMap is really a great place to harmonize GIS data.
• 22:35
  SPEAKER_01: Humans still play a very critical role there.
• 22:38
  SPEAKER_01: So that's important for us all.
• 22:42
  SPEAKER_01: Brilliant.
• 22:43
  SPEAKER_01: Okay.
• 22:43
  SPEAKER_01: And you're personally active on LinkedIn, I think.
• 22:46
  SPEAKER_01: So that's probably another place people can follow.
• 22:48
  SPEAKER_00: Yes.
• 22:51
  SPEAKER_00: Yeah, and have a look at our YouTube videos.
• 22:54
  I think they are the best way to have a quick understanding
• 22:57
  about what we're actually doing.
• 22:58
  SPEAKER_01: Perfect.
• 23:00
  SPEAKER_01: I'll place those links in the show notes.
• 23:04
  But once again,
• 23:04
  thanks for joining me today and telling me about this
• 23:06
  fascinating topic.
• 23:07
  SPEAKER_01: And I wish your project all the best in the future.
• 23:11
  SPEAKER_00: Yeah, thank you so much, Robin.