Latest / Satellite image deep learning / Insights from the SMAC earthquake detection challenge

Transcript

Insights from the SMAC earthquake detection challenge · Satellite image deep learning

• 0:05
  Hello,
• 0:06
  I'm Robin Cole,
• 0:06
  and welcome to the Satellite Image Deep Learning Podcast.
• 0:09
  SPEAKER_02: In this episode,
• 0:10
  SPEAKER_02: I caught up with Daniele Regiacampin,
• 0:12
  SPEAKER_02: the organizer of the SMAC Earthquake Detection Challenge,
• 0:16
  SPEAKER_02: and Giorgio Morales,
• 0:17
  SPEAKER_02: its winner.
• 0:18
  The challenge invited participants to leverage Sentinel-1
• 0:21
  satellite imagery to
• 0:22
  identify earthquake-affected areas and measure the strength of
• 0:25
  events by promoting
• 0:27
  SPEAKER_02: scalable and resource-efficient solutions.
• 0:30
  Giorgio shared his innovative approach that secured first place
• 0:33
  when we explored
• 0:34
  SPEAKER_02: the effort behind designing and solving such a meaningful challenge.
• 0:38
  This conversation provides valuable insights into developing
• 0:40
  effective solutions
• 0:42
  SPEAKER_02: and showcases the potential of satellite data in earthquake monitoring.
• 0:46
  SPEAKER_02: If you're watching this on YouTube,
• 0:47
  SPEAKER_02: please consider liking the video and subscribing to the channel,
• 0:50
  SPEAKER_02: as this really helps to reach a wider audience.
• 0:52
  SPEAKER_02: Thank you.
• 0:53
  SPEAKER_02: Hi, gentlemen.
• 0:54
  SPEAKER_02: Welcome to the podcast.
• 0:55
  SPEAKER_02: Daniele, great to have you back on the podcast.
• 0:57
  SPEAKER_02: And Giorgio, welcome for the first time.
• 1:00
  SPEAKER_02: Thank you for having me.
• 1:04
  SPEAKER_02: Yes, Daniele, let's begin with you.
• 1:06
  SPEAKER_02: How about you give us an overview of how the competition went?
• 1:10
  SPEAKER_01: Yeah, thanks for listening to the invitation again.
• 1:13
  SPEAKER_01: And yeah, I will try to give you an overview.
• 1:16
  So there are also some other people that have taken part other
• 1:19
  than Giorgio,
• 1:19
  SPEAKER_01: for sure,
• 1:20
  SPEAKER_01: that was our winner.
• 1:22
  SPEAKER_01: So...
• 1:24
  SPEAKER_01: In the competition,
• 1:26
  we ask the people to investigate if it's possible to understand
• 1:32
  the magnitude of
• 1:33
  the earthquakes and also if area was affected by some certain
• 1:37
  seismic activity,
• 1:39
  SPEAKER_01: just to foster.
• 1:41
  SPEAKER_01: Some people ask me in a conference, what is the purpose of this?
• 1:47
  SPEAKER_01: So my answer to this is,
• 1:49
  SPEAKER_01: With this,
• 1:50
  we can simply try to help people that are domain experts to
• 1:55
  investigate with finer
• 1:57
  SPEAKER_01: granularity the area of interest,
• 1:59
  because scrapping all around the globe for scientific activity
• 2:02
  is quite a big issue.
• 2:04
  So if we can try to understand which areas are the most
• 2:07
  interesting for the people
• 2:09
  SPEAKER_01: that are domain experts,
• 2:10
  SPEAKER_01: it's for sure
• 2:12
  a good way to improve our analysis and also improve the fastening,
• 2:18
  the procedure,
• 2:19
  I would say.
• 2:20
  SPEAKER_01: So in the competition,
• 2:22
  SPEAKER_01: we ask people to detect the magnitude and also,
• 2:24
  SPEAKER_01: as I mentioned before,
• 2:26
  SPEAKER_01: if an error was affected,
• 2:28
  SPEAKER_01: with some imagery taken from a satellite that is Sentinel-1,
• 2:33
  SPEAKER_01: so from radar imagery.
• 2:35
  SPEAKER_01: So they are quite
• 2:37
  SPEAKER_01: I would say, reliable every day, every night, in any condition.
• 2:41
  SPEAKER_01: So they can easily deploy it in any situation.
• 2:45
  SPEAKER_01: Also,
• 2:45
  SPEAKER_01: they mentioned that nowadays many people are using radar,
• 2:49
  SPEAKER_01: low-cost radar,
• 2:50
  SPEAKER_01: to monitor other things.
• 2:51
  SPEAKER_01: They are launching many of these.
• 2:52
  So they,
• 2:54
  for sure,
• 2:54
  are quite low-cost things that could be deployed and collecting data.
• 3:00
  SPEAKER_01: So we collected, I would say, 20, 30 participants, more or less.
• 3:08
  SPEAKER_01: Some people have waited during the competition because everyone is doing,
• 3:12
  SPEAKER_01: I would say,
• 3:13
  SPEAKER_01: for the sake of the competition and for curiosity,
• 3:16
  SPEAKER_01: mainly.
• 3:19
  SPEAKER_01: was like the previous year for the other competition.
• 3:24
  SPEAKER_01: So I would say the numbers for these topics,
• 3:28
  SPEAKER_01: so related to remote sensing,
• 3:30
  SPEAKER_01: observation,
• 3:31
  SPEAKER_01: crisis management,
• 3:32
  SPEAKER_01: et cetera.
• 3:34
  In this conference,
• 3:35
  I would like that this ECML are like 30 people,
• 3:38
  more or less.
• 3:39
  Then I would say that at the end,
• 3:42
  five,
• 3:42
  seven people reached,
• 3:44
  I would say,
• 3:44
  the last phase.
• 3:45
  SPEAKER_01: So they have their whole solution that could work.
• 3:51
  With respect to the previous year,
• 3:53
  what I noticed that all the people stick with my data.
• 3:58
  SPEAKER_01: The last year,
• 3:59
  SPEAKER_01: some people have tried to do some training,
• 4:02
  SPEAKER_01: use external data,
• 4:04
  SPEAKER_01: publicly available data,
• 4:05
  SPEAKER_01: for sure.
• 4:06
  SPEAKER_01: Also this year,
• 4:07
  we tried to reduce this constraint because we don't want people
• 4:10
  to use some
• 4:11
  SPEAKER_01: commercial data or
• 4:13
  other things like that to limit the reproducibility and we ask
• 4:17
  them to also share
• 4:18
  their solution with people because we have to learn at least
• 4:22
  it's a competition and
• 4:24
  SPEAKER_01: so this is what is the purpose of this i would say so
• 4:30
  SPEAKER_01: From this year,
• 4:31
  SPEAKER_01: I would say we have collected many people that are more expert in this,
• 4:35
  SPEAKER_01: like Giorgio and other participants,
• 4:37
  SPEAKER_01: worked many times with Sentinel-1,
• 4:40
  SPEAKER_01: radiometric imagery.
• 4:41
  SPEAKER_01: So they provided us with a more expertise.
• 4:45
  SPEAKER_01: I would say that I'm a more general purpose investigator,
• 4:49
  SPEAKER_01: mainly computer science oriented,
• 4:50
  SPEAKER_01: I would say.
• 4:51
  SPEAKER_01: So that's also the objective,
• 4:54
  SPEAKER_01: I would say,
• 4:54
  SPEAKER_01: these competitions also try to collect people from different expertise,
• 5:01
  SPEAKER_01: from different,
• 5:02
  SPEAKER_01: I would say,
• 5:02
  SPEAKER_01: fields also,
• 5:03
  SPEAKER_01: if possible,
• 5:04
  SPEAKER_01: and try to,
• 5:05
  SPEAKER_01: you know,
• 5:07
  gain the,
• 5:08
  collect them to investigate these possibilities because we need them.
• 5:13
  SPEAKER_01: And also it's like, you know, an emerging field.
• 5:15
  So we need people with different expertise from not only
• 5:18
  computer engineering,
• 5:19
  SPEAKER_01: from in this case,
• 5:20
  SPEAKER_01: geophysical,
• 5:21
  SPEAKER_01: I would say background.
• 5:23
  SPEAKER_01: but also as you are one of them from probably physics field, et cetera.
• 5:28
  SPEAKER_01: So that's good to have other people that are experts in other things.
• 5:34
  SPEAKER_01: So they can also try to understand if there are some issues
• 5:39
  SPEAKER_01: I would say, with the data.
• 5:40
  SPEAKER_01: But because at the first time,
• 5:42
  SPEAKER_01: when you provide the data,
• 5:43
  SPEAKER_01: when you work with the data,
• 5:45
  you try to focus on certain aspects because you don't really
• 5:48
  know exactly what you
• 5:50
  SPEAKER_01: are doing.
• 5:51
  SPEAKER_01: So it's important,
• 5:52
  SPEAKER_01: I would say,
• 5:53
  SPEAKER_01: to let other people that are not biased in some sense,
• 5:56
  SPEAKER_01: I would say,
• 5:57
  SPEAKER_01: by what you have done,
• 5:59
  SPEAKER_01: to understand exactly if there are some issues and flaws,
• 6:03
  SPEAKER_01: if there are some better way to evaluate the problem,
• 6:08
  The problem is even better solvable,
• 6:11
  like this is the point of the competition.
• 6:13
  SPEAKER_01: If there are better solution, easier solution.
• 6:16
  SPEAKER_01: There are also many other competition around,
• 6:19
  SPEAKER_01: like in other conferences,
• 6:20
  SPEAKER_01: also like other workshops that are tailored for competition,
• 6:24
  SPEAKER_01: like Semival,
• 6:24
  SPEAKER_01: for example,
• 6:26
  SPEAKER_01: or Track that is like for
• 6:28
  SPEAKER_01: 20 years right now that they do this.
• 6:29
  It's like simply providing solution that works for people and
• 6:34
  also research,
• 6:35
  SPEAKER_01: because sometimes we still focus on providing something new,
• 6:38
  SPEAKER_01: but they work probably sometimes worse than previous solution.
• 6:46
  SPEAKER_01: That's the problem sometimes.
• 6:48
  SPEAKER_01: We simply lose our focus as a researcher, I would say.
• 6:51
  SPEAKER_01: Probably in industries like different, but yeah.
• 6:54
  SPEAKER_01: That's my view, I would say, of the whole competition.
• 6:57
  SPEAKER_01: I don't know if I missed something, probably.
• 6:59
  SPEAKER_02: Thank you.
• 7:01
  SPEAKER_02: Yeah, I mean, I agree.
• 7:02
  It's a very interesting field to work in just because you have
• 7:05
  all these people
• 7:05
  SPEAKER_02: coming from different backgrounds and contributing together,
• 7:08
  SPEAKER_02: some working on these problems for the first time,
• 7:11
  SPEAKER_02: some with deep domain expertise.
• 7:13
  SPEAKER_02: On that point,
• 7:14
  it'd be great to hand over to Giorgio and hear about how you
• 7:17
  tackle this challenge.
• 7:19
  SPEAKER_00: Yeah.
• 7:20
  SPEAKER_00: Thank you for the invitation.
• 7:22
  SPEAKER_00: Thank you, Daniela, for...
• 7:24
  SPEAKER_00: organizing this competition.
• 7:26
  SPEAKER_00: Um,
• 7:27
  SPEAKER_00: probably I would like to start saying that,
• 7:30
  SPEAKER_00: um,
• 7:32
  SPEAKER_00: in this type of competitions,
• 7:34
  SPEAKER_00: when you receive the data,
• 7:36
  SPEAKER_00: uh,
• 7:37
  SPEAKER_00: one tends to,
• 7:38
  SPEAKER_00: to use basically,
• 7:41
  SPEAKER_00: uh,
• 7:42
  SPEAKER_00: a very
• 7:43
  common approach like training a neural network or a
• 7:47
  convolutional neural network in
• 7:48
  this case that hopefully is going to capture the complexity of
• 7:52
  the data or that
• 7:53
  SPEAKER_00: kind of thing.
• 7:56
  But I guess my approach was a little bit different this time
• 8:00
  because I noticed that
• 8:01
  SPEAKER_00: when the data was given,
• 8:03
  SPEAKER_00: I mean,
• 8:05
  SPEAKER_00: just to give a little bit of context,
• 8:07
  SPEAKER_00: we are given images that correspond to the pre-event and another image,
• 8:12
  SPEAKER_00: another corresponding image that corresponds to the post-event.
• 8:16
  So basically the task is to compare these two images and for
• 8:21
  the first time,
• 8:24
  challenge is to detect if there was an earthquake or not and
• 8:29
  well the second one is
• 8:29
  to calculate if there was an earthquake what was the magnitude
• 8:33
  but just for the
• 8:35
  sake of comparison uh what i started doing was to actually
• 8:39
  compare pixel per pixel
• 8:42
  and it was very uh suspiciously easy when when it comes to
• 8:48
  compare pixel to pixel
• 8:49
  SPEAKER_00: because for the case of a
• 8:53
  events in which there was no earthquake,
• 8:56
  basically the pixels were exactly the same.
• 8:58
  I mean,
• 8:59
  the backscattering coefficients for the VV image,
• 9:05
  which basically is the image that
• 9:08
  sends a signal with a vertical filter and receives the signals
• 9:14
  with the same kind of filter.
• 9:15
  SPEAKER_00: That's one band of the image.
• 9:18
  And the other one is sends the signal with vertical filter and
• 9:23
  receives it with
• 9:25
  SPEAKER_00: horizontal filter.
• 9:26
  SPEAKER_00: So we have an image with two bands.
• 9:29
  SPEAKER_00: So there was no change for at least one of them.
• 9:36
  And I think that the main reason is that the images were
• 9:39
  captured within a very
• 9:42
  SPEAKER_00: short period of time.
• 9:45
  Maybe I have to say something about that later,
• 9:47
  but basically that was the main approach.
• 9:52
  counting the number of pixels or the ratio of pixels in the
• 9:56
  image that did not
• 9:57
  SPEAKER_00: present any kind of change.
• 9:59
  SPEAKER_00: That was the main classifier.
• 10:02
  And the only hyperparameter that I decided to use was the
• 10:06
  threshold of this ratio.
• 10:08
  SPEAKER_00: So if this ratio is greater than, say, epsilon,
• 10:12
  SPEAKER_00: then we can say that there was or was not an earthquake event.
• 10:18
  SPEAKER_00: So that basically gave me or allowed me to obtain,
• 10:23
  SPEAKER_00: I think,
• 10:25
  SPEAKER_00: almost 99 of F1 score in the training set.
• 10:32
  SPEAKER_00: It was very interesting to see that there was no false negative.
• 10:38
  SPEAKER_00: There was only false positives, eight false positives in particular.
• 10:42
  SPEAKER_00: And those cases correspond to what I suspect it was images of the shore.
• 10:50
  SPEAKER_00: So basically images of the sea.
• 10:53
  SPEAKER_00: And you could see the shapes of the waves and everything.
• 10:56
  So,
• 10:56
  of course,
• 10:57
  if there was earthquake or not earthquake,
• 11:01
  there's always going to be a change.
• 11:03
  SPEAKER_00: So I think it makes sense.
• 11:07
  I guess one critique about my approach is that it's basically a
• 11:13
  change detection in general.
• 11:15
  SPEAKER_00: It works for the purpose of the competition,
• 11:17
  but it doesn't mean that we now have or are able to collect the
• 11:22
  representation of
• 11:25
  or the knowledge to discern between what is change that
• 11:29
  corresponds specifically to
• 11:31
  SPEAKER_00: earthquake or change that corresponds to fires or something like that.
• 11:36
  SPEAKER_00: So for the purpose of the competition, it works.
• 11:40
  SPEAKER_00: And that was the first task.
• 11:41
  SPEAKER_00: The second task was the magnitude estimation.
• 11:45
  And for that particular case,
• 11:48
  I tried in the beginning to use some kind of regression models.
• 11:53
  SPEAKER_00: Again,
• 11:54
  SPEAKER_00: neural network based models in the beginning.
• 11:58
  But I noticed that every different type of technique that I
• 12:02
  decided to use was
• 12:03
  SPEAKER_00: leading to the same kind of error.
• 12:05
  So basically,
• 12:06
  I combined all the information from the training set,
• 12:09
  validation set.
• 12:10
  I performed cross-validation and I noticed that it didn't
• 12:15
  matter what kind of class
• 12:17
  SPEAKER_00: or regression model I decided to use.
• 12:20
  SPEAKER_00: the error or the accuracy was not statistically significant.
• 12:26
  So one of the conclusions that I arrived is that basically we
• 12:31
  don't have enough
• 12:32
  SPEAKER_00: information to understand or estimate the magnitude of the earthquake.
• 12:38
  SPEAKER_00: So conceptually, I think that's one of the
• 12:42
  one of the results of the competition to understand that given
• 12:46
  one image from
• 12:48
  before and an image from after is not enough in itself to
• 12:53
  estimate the magnitude.
• 12:56
  SPEAKER_00: Say, for example,
• 12:58
  SPEAKER_00: If we are located at one specific point, where is the hyper center?
• 13:03
  SPEAKER_00: Is it hyper center?
• 13:04
  SPEAKER_00: Hyper center?
• 13:06
  SPEAKER_01: The one at the surface.
• 13:09
  SPEAKER_01: There is one in the bottom, inside the heart.
• 13:14
  SPEAKER_00: Okay, for my example, let's say epicenter in that case.
• 13:19
  SPEAKER_00: So let's suppose that we are located at the epicenter.
• 13:24
  The magnitude or the effects that correspond to the surrounding
• 13:29
  area are in an event,
• 13:31
  SPEAKER_00: let's say,
• 13:32
  SPEAKER_00: of...
• 13:33
  of a magnitude 7,
• 13:35
  the effects are going to be very noticeable around that area.
• 13:39
  But if we move away,
• 13:41
  let's say five kilometers,
• 13:43
  the effects are going to be different.
• 13:44
  SPEAKER_00: So if we only have the images,
• 13:47
  SPEAKER_00: without knowing in which specific point we are,
• 13:51
  SPEAKER_00: it's not going to be enough.
• 13:55
  SPEAKER_00: That's the first observation.
• 13:56
  At the minimum,
• 13:57
  we will need to know what's the distance with respect to the epicenter.
• 14:02
  SPEAKER_00: And second one,
• 14:04
  SPEAKER_00: if we repeat that same example and we imagine that we are in the desert,
• 14:11
  SPEAKER_00: the effects are going to be different because of the soil properties and,
• 14:18
  SPEAKER_00: well,
• 14:18
  SPEAKER_00: basically soil properties,
• 14:19
  SPEAKER_00: I guess.
• 14:21
  SPEAKER_00: But if we repeat the example in the mountains,
• 14:25
  SPEAKER_00: the effect is also going to be different.
• 14:28
  SPEAKER_00: So we will also need to have access to the information about soil types,
• 14:35
  SPEAKER_00: environmental conditions,
• 14:37
  SPEAKER_00: and so on.
• 14:38
  SPEAKER_00: So I guess that's something that,
• 14:43
  SPEAKER_00: as I said,
• 14:44
  SPEAKER_00: it's something that we learned from the competition.
• 14:49
  SPEAKER_00: And something else, another observation is that probably it's going to be
• 14:55
  A mistake for future approaches to try to gather or collect all
• 15:02
  the images that
• 15:02
  correspond to earthquakes from all around the world and try to
• 15:07
  train a single model.
• 15:08
  SPEAKER_00: Basically because of the same reason that I said before.
• 15:12
  And I think in the report I mentioned something similar happens
• 15:18
  in precision
• 15:19
  SPEAKER_00: agriculture because...
• 15:22
  SPEAKER_00: Normally,
• 15:23
  SPEAKER_00: when you study one crop,
• 15:26
  SPEAKER_00: let's say corn,
• 15:27
  SPEAKER_00: these crops don't behave the same in different parts of the world,
• 15:31
  SPEAKER_00: not even within the same regions.
• 15:33
  So one usually tends to train a model that is specific to each
• 15:38
  farm or each field
• 15:42
  SPEAKER_00: just to capture the conditions of the soil,
• 15:44
  SPEAKER_00: the environmental conditions and so on.
• 15:47
  SPEAKER_00: Because these...
• 15:49
  SPEAKER_00: processes are very unique of the regions and are not very generalizable.
• 15:55
  SPEAKER_00: So that's also something else to take into account.
• 16:00
  SPEAKER_00: The problem or the new challenge,
• 16:02
  SPEAKER_00: I guess,
• 16:02
  is that there's not a lot of information from different parts
• 16:08
  of the world.
• 16:09
  I mean,
• 16:10
  if we focus just in South America,
• 16:12
  in the part that is near to the Pacific region,
• 16:17
  There's not a lot of cases where you can gather information and
• 16:22
  train a very complete model.
• 16:24
  SPEAKER_00: So I guess one has to make a lot of assumptions in that case.
• 16:31
  SPEAKER_00: But again, it's a very complex problem and that's why it's interesting.
• 16:36
  SPEAKER_02: Yeah,
• 16:37
  so earthquakes being relatively rare and also very geography
• 16:43
  dependent means that
• 16:44
  SPEAKER_02: creating a large data set is always going to be really challenging.
• 16:48
  SPEAKER_02: And I guess,
• 16:48
  SPEAKER_02: Daniele,
• 16:49
  this is something that's going to go back to you when you think
• 16:52
  about taking this further.
• 16:53
  And it's kind of interesting to me that the detection side of
• 16:56
  it is really straightforward,
• 16:57
  SPEAKER_02: albeit within the kind of context of,
• 16:59
  SPEAKER_02: you know,
• 17:00
  SPEAKER_02: before and after quite limited time zones.
• 17:04
  But yeah,
• 17:05
  you did originally have plans to regress the location of the epicenter,
• 17:10
  right?
• 17:12
  SPEAKER_01: Yeah, still, I'm working with this right now.
• 17:15
  SPEAKER_01: I will say in these months, I will say.
• 17:18
  So I think in this case,
• 17:20
  it's different from what I've seen and from what you can see.
• 17:24
  SPEAKER_01: Because with this type of product, it's like ground-range detected.
• 17:28
  SPEAKER_01: So yeah,
• 17:29
  SPEAKER_01: as you mentioned before,
• 17:30
  SPEAKER_01: by George,
• 17:31
  SPEAKER_01: it's quite strange because it's ground-range detected.
• 17:34
  But some samples are literally quite in the sea,
• 17:39
  but probably because there is some sort of
• 17:42
  SPEAKER_01: little land.
• 17:43
  And so the satellite says,
• 17:45
  yeah,
• 17:45
  OK,
• 17:45
  there is some ground,
• 17:47
  but that's not true.
• 17:48
  So when you collect them,
• 17:50
  you simply trust the Copernicus,
• 17:53
  I would say,
• 17:54
  system.
• 17:56
  SPEAKER_01: And yes, if they tell you that these
• 17:59
  we say product is working with ground,
• 18:01
  you stick with,
• 18:02
  OK,
• 18:03
  it's ground in some way.
• 18:04
  SPEAKER_01: But yeah, for sure, it's not ground in this case.
• 18:06
  SPEAKER_01: So you can clean the collection, I would say.
• 18:08
  There are also some samples on the ground,
• 18:13
  because not everything is on the sea.
• 18:16
  SPEAKER_01: So now I'll try to work with some other type of data of Sentinel one.
• 18:22
  SPEAKER_01: It's like SLC, so the complex one.
• 18:25
  So in this case,
• 18:26
  you can simply try to process them to get the interferogram.
• 18:30
  SPEAKER_01: So people stick with interferogram right now.
• 18:34
  SPEAKER_01: And generally, you can see they provide some interference.
• 18:40
  So you can try to understand at least in a different way,
• 18:44
  because the problem is that.
• 18:46
  As you mentioned probably also by Giorgio,
• 18:48
  is that this type of data,
• 18:50
  like GRD,
• 18:53
  SPEAKER_01: As not,
• 18:54
  SPEAKER_01: I would say,
• 18:55
  SPEAKER_01: many things that are not independent,
• 18:58
  SPEAKER_01: I would say,
• 18:58
  SPEAKER_01: of the place and other things like that.
• 19:01
  SPEAKER_01: Well,
• 19:02
  SPEAKER_01: for example,
• 19:02
  if you search for SLC and also if you search for people that
• 19:06
  are doing this,
• 19:07
  SPEAKER_01: I would say,
• 19:08
  SPEAKER_01: by hand,
• 19:09
  SPEAKER_01: they get,
• 19:10
  SPEAKER_01: I would say,
• 19:10
  SPEAKER_01: like sort of cleaned
• 19:13
  SPEAKER_01: I would say, phase map.
• 19:16
  So you get,
• 19:16
  imagine,
• 19:17
  a simply matrix with the information about the phase in each point.
• 19:21
  SPEAKER_01: So it's different.
• 19:23
  I don't really know,
• 19:24
  also in this case,
• 19:25
  as before,
• 19:26
  if some sort of model could get something.
• 19:30
  SPEAKER_01: But at least you can see, I would say, in a large area,
• 19:35
  SPEAKER_01: where an earthquake happened,
• 19:36
  SPEAKER_01: if there are some circles or there are some sort of,
• 19:40
  SPEAKER_01: like I would say,
• 19:41
  SPEAKER_01: some sort of walls or interference.
• 19:43
  SPEAKER_01: So you can try to do more,
• 19:47
  SPEAKER_01: a different analysis is quite agnostic since in this case,
• 19:49
  SPEAKER_01: you simply plot instead of,
• 19:52
  SPEAKER_01: you know,
• 19:53
  SPEAKER_01: the value on the range of the before and the after,
• 19:57
  SPEAKER_01: you have a phase information that in that specific,
• 20:02
  SPEAKER_01: I would say,
• 20:02
  SPEAKER_01: products,
• 20:03
  SPEAKER_01: the GRD,
• 20:04
  SPEAKER_01: you simply lose because they kept only the amplitude.
• 20:07
  SPEAKER_01: So you have another information that's probably also something that,
• 20:11
  SPEAKER_01: as Giorgio mentioned before,
• 20:13
  SPEAKER_01: could be useful also because
• 20:16
  SPEAKER_01: Probably there are some missing information.
• 20:19
  SPEAKER_01: I have tried also to include other information,
• 20:22
  SPEAKER_01: like tectonic plates and also the soil information.
• 20:29
  SPEAKER_01: But still, I do not get nothing better, I would say, after this insight.
• 20:35
  SPEAKER_01: Because the point is now,
• 20:37
  understand how you can process these type of data that are
• 20:41
  quite complex also for humans.
• 20:43
  SPEAKER_01: And also, the problem that I've also noticed is like,
• 20:47
  SPEAKER_01: the inaccuracies.
• 20:49
  SPEAKER_01: Geophysics is quite complex.
• 20:51
  SPEAKER_01: So also the information that you have about the epicenter,
• 20:56
  SPEAKER_01: the magnitude,
• 20:58
  SPEAKER_01: et cetera,
• 20:59
  SPEAKER_01: have sometimes quite a large error,
• 21:02
  SPEAKER_01: I would say.
• 21:02
  SPEAKER_01: Depends on the event, I would say.
• 21:05
  So it's sometimes difficult also to say,
• 21:07
  yeah,
• 21:08
  you have to regress this point.
• 21:09
  SPEAKER_01: You have to regress, I would say, this strength.
• 21:13
  SPEAKER_01: It's not easy.
• 21:14
  SPEAKER_01: Also, if you search for
• 21:17
  different companies or different institutions that do this type
• 21:21
  of annotation,
• 21:22
  SPEAKER_01: they don't agree.
• 21:23
  SPEAKER_01: Literally, they don't agree easily.
• 21:26
  SPEAKER_01: So that's difficult,
• 21:27
  SPEAKER_01: also,
• 21:27
  SPEAKER_01: in this case,
• 21:28
  SPEAKER_01: because I've seen other people right now that are working on this,
• 21:32
  SPEAKER_01: like damaged buildings for earthquakes,
• 21:37
  SPEAKER_01: et cetera.
• 21:38
  SPEAKER_01: There are,
• 21:39
  SPEAKER_01: I would say,
• 21:39
  SPEAKER_01: many things that you cannot do anything for earthquakes to prevent them.
• 21:44
  SPEAKER_01: They simply happen.
• 21:45
  You can try to simply understand how to recover and how to make
• 21:51
  people simply leave the places.
• 21:53
  So yeah,
• 21:54
  for sure,
• 21:54
  I would say,
• 21:55
  as Giorgio mentioned,
• 21:56
  it's an interesting way to
• 21:59
  SPEAKER_01: investigate how we can help, I would say, in this case.
• 22:02
  SPEAKER_01: Because,
• 22:02
  SPEAKER_01: yeah,
• 22:02
  SPEAKER_01: you cannot prevent,
• 22:03
  SPEAKER_01: like,
• 22:03
  SPEAKER_01: you know,
• 22:04
  SPEAKER_01: in the case of welfare,
• 22:06
  SPEAKER_01: you can simply try to prevent the welfare.
• 22:07
  SPEAKER_01: But in this case, you cannot prevent an earthquake.
• 22:09
  SPEAKER_01: They happen.
• 22:10
  SPEAKER_01: So you have to do something, I would say.
• 22:12
  SPEAKER_01: You cannot simply say,
• 22:13
  SPEAKER_01: as people pay attention to,
• 22:16
  SPEAKER_01: do not leave,
• 22:16
  SPEAKER_01: I would say,
• 22:18
  SPEAKER_01: something in the forest.
• 22:20
  So,
• 22:21
  I mean,
• 22:22
  what do you think is next for the solution that Georgia has developed?
• 22:27
  SPEAKER_01: So, yeah.
• 22:28
  SPEAKER_01: From my side,
• 22:29
  SPEAKER_01: also from what other people have done,
• 22:31
  SPEAKER_01: I would say that is interesting,
• 22:33
  SPEAKER_01: as I mentioned before,
• 22:35
  SPEAKER_01: to know that things.
• 22:36
  So I hope that people from,
• 22:38
  I will say,
• 22:39
  I provided all the data that they have.
• 22:41
  SPEAKER_01: So where they happen, the endpoints,
• 22:45
  SPEAKER_01: when, and also the data itself.
• 22:48
  So you can try to simply subsample them,
• 22:51
  try to collect a time series,
• 22:52
  try to do other things.
• 22:55
  So like I've done,
• 22:56
  say,
• 22:57
  try to collect some information from tectonics,
• 22:59
  from other information.
• 23:04
  SPEAKER_01: Yeah,
• 23:05
  SPEAKER_01: as Giorgio mentioned,
• 23:06
  SPEAKER_01: probably also,
• 23:07
  SPEAKER_01: I would say,
• 23:08
  SPEAKER_01: a first evaluation that could be done that is,
• 23:11
  SPEAKER_01: I would say,
• 23:12
  more easily understandable is if we can really detect if an
• 23:17
  area was affected by a
• 23:18
  SPEAKER_01: change or an earthquake,
• 23:20
  SPEAKER_01: because they are different.
• 23:22
  SPEAKER_01: As Giorgio mentioned,
• 23:23
  SPEAKER_01: the first thing is simply aggregate some data sets from a wildfire,
• 23:29
  SPEAKER_01: for example,
• 23:29
  SPEAKER_01: or from harder changes.
• 23:32
  SPEAKER_01: I would say cropped the changes.
• 23:35
  So and try to understand if you could discriminate between
• 23:38
  changes due to a shake
• 23:40
  of the ground or a simply wildfire that burned everything or
• 23:45
  some other event.
• 23:46
  SPEAKER_01: That could be easily done, I would say.
• 23:48
  SPEAKER_01: Because in this case, in this evaluation, I would say it's not complete.
• 23:52
  SPEAKER_01: Because,
• 23:53
  SPEAKER_01: yeah,
• 23:53
  SPEAKER_01: you have some non-airquakes,
• 23:55
  SPEAKER_01: some earthquakes,
• 23:55
  SPEAKER_01: but probably it could be more interesting for people to know if we could
• 23:59
  SPEAKER_01: discriminate these type of changes.
• 24:02
  I mean,
• 24:02
  this is more feasible,
• 24:04
  generally speaking,
• 24:04
  without much effort,
• 24:06
  I would say.
• 24:06
  If you don't get,
• 24:08
  I would say,
• 24:09
  any results,
• 24:10
  probably,
• 24:11
  it's more difficult than you can talk.
• 24:15
  SPEAKER_01: But at least it's better to give it a shot.
• 24:19
  SPEAKER_02: So it sounds like we can be looking forward to a next generation of this
• 24:22
  SPEAKER_02: competition focused on that particular task,
• 24:25
  SPEAKER_02: bringing in other metadata and other sources of imagery,
• 24:28
  SPEAKER_02: potentially.
• 24:30
  SPEAKER_02: To close,
• 24:30
  SPEAKER_02: I'd like to hand it over to you,
• 24:31
  SPEAKER_02: Giorgio,
• 24:32
  SPEAKER_02: just to hear a bit more about what you learned during this competition.
• 24:37
  I guess it's related to what Daniel said about discriminating
• 24:45
  between changes.
• 24:48
  I guess I learned that it is not very easy to see the patterns
• 24:54
  of an earthquake.
• 24:55
  SPEAKER_00: I mean,
• 24:55
  SPEAKER_00: when one imagines how an earthquake propagates,
• 25:00
  SPEAKER_00: that's usually related to what's happening underground.
• 25:05
  SPEAKER_00: some kind of lines or waves that move in circles, something like that.
• 25:09
  SPEAKER_00: But it really depends on the type of,
• 25:13
  SPEAKER_00: as I said,
• 25:13
  SPEAKER_00: the type of soil,
• 25:14
  SPEAKER_00: but also if it's an urban area that is covering asphalt,
• 25:19
  SPEAKER_00: you can't directly see,
• 25:20
  SPEAKER_00: even if there are effects or changes,
• 25:23
  SPEAKER_00: you can't see them.
• 25:25
  So related to what Daniela just said about first classifying
• 25:29
  change detections that
• 25:31
  SPEAKER_00: correspond to crop changes or
• 25:35
  SPEAKER_00: urban changes or something like that,
• 25:37
  what I imagined that they would be feasible would be like a
• 25:42
  hierarchical classification.
• 25:44
  SPEAKER_00: So first,
• 25:46
  SPEAKER_00: the easy type of detection would be crop changes or something like that,
• 25:52
  SPEAKER_00: then move to fires,
• 25:53
  SPEAKER_00: move to that kind of thing.
• 25:57
  classify all those changes with some kind of uncertainty
• 26:01
  prediction intervals,
• 26:02
  for example.
• 26:03
  SPEAKER_00: That will be interesting to see.
• 26:05
  SPEAKER_00: And having a map with these
• 26:07
  different type of changes in different regions of a city or an
• 26:15
  area will be
• 26:16
  SPEAKER_00: something very useful to give to the authorities,
• 26:18
  SPEAKER_00: for example,
• 26:19
  SPEAKER_00: to plan accordingly.
• 26:22
  SPEAKER_00: I guess that will be interesting.
• 26:24
  SPEAKER_00: It's not exactly my research area,
• 26:30
  SPEAKER_00: but I think what I've done in the past related to remote sensing
• 26:36
  SPEAKER_00: makes me think that that's possible.
• 26:38
  SPEAKER_00: And again, interesting.
• 26:40
  SPEAKER_00: It's an open problem and I'm all about open problems, I guess.
• 26:45
  SPEAKER_02: Fantastic.
• 26:45
  Well,
• 26:45
  thank you for sharing your insights and your solution to the competition.
• 26:50
  SPEAKER_02: I agree.
• 26:51
  SPEAKER_02: I think that there's very exciting potential for these technologies,
• 26:54
  and hopefully this will feed through into early warning systems
• 26:58
  and greater
• 26:59
  SPEAKER_02: insights for governments and local authorities.
• 27:03
  SPEAKER_02: Once again,
• 27:03
  SPEAKER_02: thank you both for joining me,
• 27:04
  SPEAKER_02: and I look forward to hearing about the next competition.