Latest / Satellite image deep learning / Building Open GeoAI for Humanitarian Response

Transcript

Building Open GeoAI for Humanitarian Response · Satellite image deep learning

• 0:00
  [music]
• 0:06
  SPEAKER_00: Hi, Petra.
• 0:06
  SPEAKER_00: Hi, Kitsch.
• 0:07
  SPEAKER_00: Welcome to the podcast.
• 0:10
  SPEAKER_02: Hi.
• 0:10
  SPEAKER_02: Hi, Robin.
• 0:11
  SPEAKER_02: Thanks for having us.
• 0:13
  SPEAKER_00: My pleasure.
• 0:14
  SPEAKER_00: Hello.
• 0:14
  SPEAKER_00: All right.
• 0:16
  SPEAKER_00: Let's begin with some intros.
• 0:18
  SPEAKER_00: Petra, maybe you'd like to kick things off.
• 0:21
  SPEAKER_02: Great.
• 0:21
  SPEAKER_02: Hi, everyone.
• 0:22
  SPEAKER_02: Happy to be here.
• 0:24
  SPEAKER_02: My name is Petra Kangalova,
• 0:26
  SPEAKER_02: and I'm working at HOT,
• 0:27
  SPEAKER_02: humanitarian OpenStreetMap team that you'll hear more about as our tech partnership
• 0:32
  SPEAKER_02: and engagement lead.
• 0:33
  SPEAKER_02: So what my role is, working with a lot of our product owners.
• 0:37
  SPEAKER_02: You'll hear from Kshetish more engaging with our users,
• 0:41
  SPEAKER_02: working with our open source contributors,
• 0:44
  SPEAKER_02: forming partnerships, so wearing many hats, I would say.
• 0:48
  SPEAKER_02: Welcome to you, Kshetish.
• 0:54
  SPEAKER_01: Hi, Robin.
• 0:55
  SPEAKER_01: I'm Kshetish.
• 0:57
  SPEAKER_01: I'm with HOT.
• 0:58
  SPEAKER_01: As an AI engineer, I oversee AI products And the development all across HOT.
• 1:06
  SPEAKER_01: Before AI,
• 1:07
  SPEAKER_01: I was to work with data pipelines,
• 1:10
  SPEAKER_01: map data access products,
• 1:13
  SPEAKER_01: ensuring that data outputs also get started with disaster responses
• 1:19
  SPEAKER_01: and preparedness.
• 1:23
  SPEAKER_01: Fantastic.
• 1:24
  SPEAKER_01: We both belong to, there's a small team within the HOT tech team.
• 1:30
  SPEAKER_00: Fantastic.
• 1:31
  SPEAKER_00: Thanks for the intros.
• 1:32
  SPEAKER_00: Petra, maybe you'd like to tell us a bit more about OSM.
• 1:34
  SPEAKER_00: I'm particularly curious to understand how it relates to OSM as well.
• 1:40
  SPEAKER_02: Amazing.
• 1:40
  SPEAKER_02: Yeah, that's a good question that we always get asked.
• 1:42
  SPEAKER_02: So humanitarian OpenStreetMap team is an NGO.
• 1:46
  SPEAKER_02: So we're an NGO that started 15 years ago.
• 1:50
  SPEAKER_02: And we operate in four regions of the world.
• 1:52
  SPEAKER_02: So that's one of the distinction.
• 1:53
  SPEAKER_02: We focus on Asian Pacific,
• 1:55
  SPEAKER_02: Eastern Southern Africa, West and Northern Africa, Latin America and the Caribbean.
• 2:01
  SPEAKER_02: So what we focus on is filling in a gap in map data,
• 2:03
  SPEAKER_02: specifically in these regions,
• 2:05
  SPEAKER_02: and that includes generating OpenStreetMap data,
• 2:08
  SPEAKER_02: but also we use other data collection methods.
• 2:13
  SPEAKER_02: So for us,
• 2:14
  SPEAKER_02: the big component is really community-driven,
• 2:17
  SPEAKER_02: locally generated data that is used for decision-making.
• 2:22
  SPEAKER_02: What I mean by decision-making,
• 2:24
  SPEAKER_02: we started very much in the disaster response,
• 2:27
  SPEAKER_02: which is still a lot of the activities that we do with the mapping efforts,
• 2:31
  SPEAKER_02: like more recently, Colombia, Venezuela earthquake, Nepal floods, and so forth.
• 2:37
  SPEAKER_02: But a lot of the mapping work also focuses on anticipated reactions,
• 2:43
  SPEAKER_02: so disaster risk reduction, climate change, health, so it has a wider application.
• 2:49
  SPEAKER_02: And yeah,
• 2:50
  SPEAKER_02: as an organization,
• 2:51
  SPEAKER_02: we're around 70 people,
• 2:53
  SPEAKER_02: but everything we do is done in partnership with Partners,
• 2:57
  SPEAKER_02: local partners in those regions and the countries.
• 3:00
  SPEAKER_02: And yeah,
• 3:01
  SPEAKER_02: maybe I'll just say in one sentence,
• 3:02
  SPEAKER_02: I really think HOT is the core of it is the people.
• 3:06
  SPEAKER_02: So we talk about the technology,
• 3:07
  SPEAKER_02: but without the people and the communities and the knowledge that they bring,
• 3:12
  SPEAKER_02: it will not be possible.
• 3:15
  SPEAKER_00: Fantastic.
• 3:16
  SPEAKER_00: You've mentioned the areas you operate.
• 3:18
  SPEAKER_00: Maybe you'd like to briefly talk about some of the initiatives that you're engaged
• 3:22
  SPEAKER_00: in, particularly those that relate to the GeoAI aspects.
• 3:26
  SPEAKER_02: Yeah, so there are many different,
• 3:30
  SPEAKER_02: I would say, projects or initiatives happening in these areas.
• 3:37
  SPEAKER_02: So there's, for instance, one project called Know Your City.
• 3:40
  SPEAKER_02: We're focusing on mapping informal settlements in the regions, like Sierra Leone, Liberia.
• 3:48
  SPEAKER_02: More recently in Indonesia,
• 3:51
  SPEAKER_02: actually we were involved in another project in Indonesia working with
• 3:57
  SPEAKER_02: the Balinese Disaster Response Agency and really focusing on
• 4:02
  SPEAKER_02: the mapping activities were on Volcano evacuation routes,
• 4:07
  SPEAKER_02: actually, and also kind of mapping the risk of tsunamis.
• 4:11
  SPEAKER_02: So we were also involved.
• 4:12
  SPEAKER_02: So the technology there,
• 4:14
  SPEAKER_02: or maybe thinking of the GUI is,
• 4:16
  SPEAKER_02: first of all,
• 4:17
  SPEAKER_02: like we,
• 4:18
  SPEAKER_02: some of the projects are using drone imagery,
• 4:22
  SPEAKER_02: which actually open imagery that's then forming the basis of Using GeoAI,
• 4:28
  SPEAKER_02: using that open imagery to also generate the models.
• 4:33
  SPEAKER_02: I would encourage the audience to look into the website because there's
• 4:35
  SPEAKER_02: so many projects with a lot of our colleagues working on them.
• 4:39
  SPEAKER_02: So yeah, those are maybe just two examples.
• 4:43
  SPEAKER_00: Thank you for that.
• 4:44
  SPEAKER_00: It's been a very exciting place to work with all these projects taking place.
• 4:47
  SPEAKER_00: So we're going to get over a particular project which is related to models
• 4:52
  SPEAKER_00: for disaster response.
• 4:53
  SPEAKER_00: Maybe you'd like to talk a bit more about that one.
• 4:58
  SPEAKER_02: So yeah, I can start and maybe Krzysztof can pick up on this one.
• 5:02
  SPEAKER_02: I think especially more recently,
• 5:05
  SPEAKER_02: I'll start with the response to the earthquake in Venezuela,
• 5:09
  SPEAKER_02: is where Krzysztof,
• 5:12
  SPEAKER_02: without more on the specifics,
• 5:13
  SPEAKER_02: So when we talk about using J-AI in this scenario,
• 5:16
  SPEAKER_02: it's around damage assessment in the focused areas.
• 5:34
  SPEAKER_00: So there's a disaster and people need to rapidly update maps,
• 5:39
  SPEAKER_00: find out which buildings are damaged,
• 5:41
  SPEAKER_00: where emergency responders need to target, that kind of thing.
• 5:45
  SPEAKER_02: It's kind of pre-disaster,
• 5:47
  SPEAKER_02: maybe in one sense I'll explain and please come in with the more technical elements.
• 5:51
  SPEAKER_02: But obviously pre-disaster,
• 5:53
  SPEAKER_02: we have imagery that's being used,
• 5:55
  SPEAKER_02: which you can also use the models to understand the density, right?
• 5:58
  SPEAKER_02: So you can use AI to understand where the buildings were clustered.
• 6:02
  SPEAKER_02: And then with post-disaster imagery, actually, that's where the human validation step is used.
• 6:08
  SPEAKER_02: And we use an application called MapSwipe where people can look in through.
• 6:12
  SPEAKER_02: So it's called MapSwipe because it does use that logic of just looking through the imagery.
• 6:17
  SPEAKER_02: But it's people who actually check and are able to say if there is damage or not and
• 6:22
  SPEAKER_02: then generate again data from there.
• 6:25
  SPEAKER_01: But yeah, feel free
• 6:27
  SPEAKER_02: to add more.
• 6:27
  SPEAKER_02: Yeah.
• 6:30
  SPEAKER_01: Yes, I think it also connects with the overall approach, the way we work.
• 6:35
  SPEAKER_01: So it always starts with manual mapping with a lot of volunteers that are helping
• 6:43
  SPEAKER_01: and assisting in order to respond to disasters.
• 6:45
  SPEAKER_01: And once that manual mapping is finished in some part Then we take that in order
• 6:52
  SPEAKER_01: to fine-tune the AI models that we have so that it could be localized in
• 6:56
  SPEAKER_01: that specific disaster in the context and it would help us deploy that
• 7:04
  SPEAKER_01: for large areas like Petya mentioned and we can quickly get the density,
• 7:08
  SPEAKER_01: we can quickly get the hotspot of the area where we can focus Next,
• 7:13
  SPEAKER_01: that will also help us as a heart to focus in a new area
• 7:17
  SPEAKER_01: where mapping might be needed and also help responders and other people
• 7:23
  SPEAKER_01: to find these priority areas and this connects with the The workflow
• 7:28
  SPEAKER_01: that we are working on would be how do you validate that output,
• 7:33
  SPEAKER_01: not only as a HOT,
• 7:34
  SPEAKER_01: like a HOT uses different other AI models that are available out there,
• 7:39
  SPEAKER_01: and in order to validate and then see the output, okay,
• 7:43
  SPEAKER_01: is this good,
• 7:44
  SPEAKER_01: is this bad,
• 7:44
  SPEAKER_01: maybe remove some of the false positives,
• 7:47
  SPEAKER_01: and then try to clean up that data so it can be used in larger scales
• 7:51
  SPEAKER_01: for other people.
• 7:53
  SPEAKER_01: Fantastic.
• 7:57
  SPEAKER_00: It's very much driven by the people that are responding to these incidents,
• 8:02
  SPEAKER_00: basically collecting truth data on the ground that's specific to those geographies.
• 8:07
  SPEAKER_00: Over time,
• 8:08
  SPEAKER_00: you're then putting that into a pipeline so we can automate aspects of
• 8:13
  SPEAKER_00: that mapping effort.
• 8:16
  SPEAKER_01: Indeed, yes.
• 8:18
  SPEAKER_01: And that also further connects with validation.
• 8:22
  SPEAKER_01: Validation happens in both in remote and field validation.
• 8:26
  SPEAKER_01: So HOT has a community.
• 8:29
  SPEAKER_01: We have a tool with a chat map.
• 8:31
  SPEAKER_01: People also use it.
• 8:31
  SPEAKER_01: These go with the WhatsApp on their phone,
• 8:35
  SPEAKER_01: try to see, okay, yes, this is damaged, this is not damaged.
• 8:39
  SPEAKER_01: And then there are people who are on their phone,
• 8:43
  SPEAKER_01: but looking at the remote imagery with the AI predictions on their phone.
• 8:48
  SPEAKER_01: Sometimes it's AI predictions,
• 8:50
  SPEAKER_01: sometimes it's some annual data sets,
• 8:53
  SPEAKER_01: and then they would validate and they would try to find the areas to generate
• 8:59
  SPEAKER_01: a better output.
• 9:02
  SPEAKER_01: Fantastic.
• 9:04
  SPEAKER_00: And so right now you have a call.
• 9:06
  SPEAKER_00: You're looking for the community to contribute AI models.
• 9:10
  SPEAKER_00: Is that accurate?
• 9:13
  SPEAKER_01: Yes, that's accurate.
• 9:15
  SPEAKER_01: I think PJ can also mention more about the call,
• 9:18
  SPEAKER_01: and this is where I think is useful for all other GUAI researchers to contribute
• 9:26
  SPEAKER_01: and join the mission.
• 9:27
  SPEAKER_01: It's hard work in open source ecosystem.
• 9:30
  SPEAKER_01: We value about the data being open,
• 9:34
  SPEAKER_01: model being open,
• 9:35
  SPEAKER_01: and also the Other researchers and developers working in the domain.
• 9:40
  SPEAKER_01: And since this will also connect with, I think, the next point was how do we make that happen?
• 9:47
  SPEAKER_01: How do we make sure that at least we can have different,
• 9:52
  SPEAKER_01: because AI models in every six months,
• 9:54
  SPEAKER_01: you will see another model coming into the market,
• 9:57
  SPEAKER_01: or people are doing excellent researches around domain,
• 10:00
  SPEAKER_01: and how do we make sure to use them?
• 10:02
  SPEAKER_01: How do they produce the output for the data?
• 10:07
  SPEAKER_01: We thought, okay,
• 10:08
  SPEAKER_01: maybe since it is changing so fast,
• 10:12
  SPEAKER_01: and for us,
• 10:12
  SPEAKER_01: we need to adapt,
• 10:13
  SPEAKER_01: and people need to know,
• 10:14
  SPEAKER_01: like, there is a new model in this area, let me check how it works for our area.
• 10:19
  SPEAKER_01: Try to fine-tune it with the data that we have,
• 10:22
  SPEAKER_01: or we collect in each new disasters,
• 10:25
  SPEAKER_01: and then we thought, let's invite all other UAI researchers and
• 10:33
  SPEAKER_01: So
• 10:40
  SPEAKER_00: the community will contribute models for specific kinds of tasks and
• 10:44
  SPEAKER_00: the intention is that over time they get updated as
• 10:47
  SPEAKER_00: the organization gets more truth data and that becomes more useful in the field.
• 10:53
  SPEAKER_01: Indeed, and same for the community as well.
• 10:56
  SPEAKER_01: So people can get training data sets who contributed the model.
• 11:01
  SPEAKER_01: They would see how their model is being used.
• 11:03
  SPEAKER_01: People are like how they are getting fine-tuned.
• 11:06
  SPEAKER_01: And also over time,
• 11:09
  SPEAKER_01: I would imagine that training data would grow bigger,
• 11:14
  SPEAKER_01: especially in the area that we work on, Global South and the low-income countries.
• 11:22
  SPEAKER_01: It's also, I would say, a two-way process for open source communities.
• 11:29
  SPEAKER_02: I was thinking as you were speaking about the call.
• 11:34
  SPEAKER_02: I didn't say that at the beginning, but I'm not a developer or data scientist.
• 11:39
  SPEAKER_02: I've worked in tech and I understand it, but I always say I'm kind of the bridge.
• 11:44
  SPEAKER_02: The two calls that we talked about,
• 11:46
  SPEAKER_02: one is for more developers,
• 11:48
  SPEAKER_02: the GIA models,
• 11:49
  SPEAKER_02: but we also had this Map Your City with FAIR, which is more the community call.
• 11:53
  SPEAKER_02: So this is really about communities generating the training data sets.
• 11:57
  SPEAKER_02: So that's what Kshitish talked about.
• 11:59
  SPEAKER_02: And I think that's really important,
• 12:01
  SPEAKER_02: even people who are not so familiar with AI,
• 12:03
  SPEAKER_02: but the training data,
• 12:05
  SPEAKER_02: you know,
• 12:05
  SPEAKER_02: in the regions where we work,
• 12:07
  SPEAKER_02: there isn't sufficient training data,
• 12:09
  SPEAKER_02: it's not relevant, then the models are not working.
• 12:13
  SPEAKER_02: So that call is something that we are personally really excited to see.
• 12:18
  SPEAKER_02: And when we talk about mapping or training data,
• 12:21
  SPEAKER_02: you know,
• 12:21
  SPEAKER_02: here we're talking about people looking at like rooftops,
• 12:24
  SPEAKER_02: looking at like Water tanks, you know, in this area.
• 12:28
  SPEAKER_02: So it's generating relevant data or we have quite a few like applications as well
• 12:33
  SPEAKER_02: that are more connected to like waste, you know, waste management.
• 12:39
  SPEAKER_02: That's also critical,
• 12:41
  SPEAKER_02: but it's, yeah, I think with this one is really generating the data locally.
• 12:46
  SPEAKER_02: And then in that way,
• 12:47
  SPEAKER_02: the models can be fine-tuned and in the regions where we operate,
• 12:51
  SPEAKER_02: AI can really, you know, add value.
• 12:53
  SPEAKER_02: Like when we talk about like where,
• 12:55
  SPEAKER_02: you know, it's not AI for the sake of it, but actually bringing, yeah, adding value.
• 13:03
  SPEAKER_00: Yeah,
• 13:03
  SPEAKER_00: and I think it's that combined effort of bringing the machine learning experts and
• 13:08
  SPEAKER_00: the people creating the data who understand the region together,
• 13:11
  SPEAKER_00: that's really going to result in high quality solutions.
• 13:18
  SPEAKER_01: Indeed, indeed.
• 13:19
  SPEAKER_01: And there are two separate calls currently which are open.
• 13:25
  SPEAKER_01: Well, I think one recently closed for the generating training data and
• 13:29
  SPEAKER_01: the project will start now and they will start collecting training data
• 13:34
  SPEAKER_01: in different parts of the world.
• 13:36
  SPEAKER_01: So far, there are more than 25 plus projects that are happening.
• 13:41
  SPEAKER_01: And the call for the contributing GAI model size is still open,
• 13:47
  SPEAKER_01: which is inviting the researchers and developers to contribute their GAI models.
• 13:52
  SPEAKER_00: Fantastic.
• 13:54
  SPEAKER_00: I've had a quick look at the website.
• 13:55
  SPEAKER_00: It looks like you're looking for maybe seven specific kinds of models.
• 13:59
  SPEAKER_00: Is that right?
• 14:02
  SPEAKER_01: Yes, the list is pretty open.
• 14:04
  SPEAKER_01: We put our priorities and the areas that we would like to focus on.
• 14:08
  SPEAKER_01: However, it's not limited into those sets of the themes that we have identified.
• 14:15
  SPEAKER_01: Of course, if there are more use cases, they are welcome.
• 14:18
  SPEAKER_01: But those are our,
• 14:19
  SPEAKER_01: like I would say, priorities and the specific areas that we identified initially.
• 14:24
  SPEAKER_00: Right.
• 14:27
  SPEAKER_00: And these are tasks that I think many in the Aerial community will be familiar
• 14:31
  SPEAKER_00: with object detection,
• 14:33
  SPEAKER_00: segmentation, and obviously related to things like structures and environment, etc.
• 14:41
  SPEAKER_02: Yeah, no, exactly.
• 14:42
  SPEAKER_02: And I think I was going to add here,
• 14:43
  SPEAKER_02: actually,
• 14:44
  SPEAKER_02: some of them are also,
• 14:44
  SPEAKER_02: you know,
• 14:44
  SPEAKER_02: we have so many projects,
• 14:45
  SPEAKER_02: I didn't use all the examples,
• 14:47
  SPEAKER_02: but for instance,
• 14:47
  SPEAKER_02: Seagrass,
• 14:48
  SPEAKER_02: you know,
• 14:48
  SPEAKER_02: they are real projects with real needs that are That are happening
• 14:53
  SPEAKER_02: and Podstaff are collaborating.
• 14:55
  SPEAKER_02: So I think the proposal of what should be,
• 14:59
  SPEAKER_02: the model should focus on is,
• 15:01
  SPEAKER_02: yeah, it's coming from the needs of the organizations and people we work with.
• 15:08
  SPEAKER_00: Of course.
• 15:09
  SPEAKER_00: And it's quite interesting,
• 15:09
  SPEAKER_00: the page lists also the expectations around the models themselves,
• 15:13
  SPEAKER_00: about incorporating open code,
• 15:16
  SPEAKER_00: for instance, and how the weights themselves should be distributed.
• 15:21
  SPEAKER_00: Could you comment on that a bit more?
• 15:25
  SPEAKER_01: Yeah, that's, I can comment on that.
• 15:28
  SPEAKER_01: So, for example,
• 15:29
  SPEAKER_01: we identified it as different steps of the pipeline,
• 15:34
  SPEAKER_01: like let's say,
• 15:35
  SPEAKER_01: as we know the way we are,
• 15:38
  SPEAKER_01: if we are inviting different Collaborate researchers and developers,
• 15:41
  SPEAKER_01: all of them will have own methods and own ways to do pre-processing,
• 15:45
  SPEAKER_01: post-processing, and the models that they developed.
• 15:48
  SPEAKER_01: That's why it is divided into four separate pipelines,
• 15:52
  SPEAKER_01: so they are open to To define pre-processing code,
• 15:57
  SPEAKER_01: their inference,
• 15:57
  SPEAKER_01: how they would like to interpret the results of that model,
• 16:01
  SPEAKER_01: what kind of post-processing techniques that they use,
• 16:04
  SPEAKER_01: and how they recommend people to do the fine-tuning.
• 16:07
  SPEAKER_01: Some models require blacktop mode training data,
• 16:09
  SPEAKER_01: some models require, okay, well,
• 16:11
  SPEAKER_01: the small amount of chips and labels are enough,
• 16:13
  SPEAKER_01: but they would like to label it in a specific way.
• 16:16
  SPEAKER_01: Hence, those pipelines will serve as kind of contract for that model.
• 16:21
  SPEAKER_01: And it should be open source licensed.
• 16:26
  SPEAKER_01: So there are different couple of licenses that we support.
• 16:31
  SPEAKER_01: And they are also listed in the website as well.
• 16:33
  SPEAKER_01: And the modal source,
• 16:36
  SPEAKER_01: whenever there is a modal word,
• 16:37
  SPEAKER_01: it should be distributed so that it can be used by any other people.
• 16:42
  SPEAKER_01: And it should be described in a stack format.
• 16:44
  SPEAKER_01: I think many of us,
• 16:46
  SPEAKER_01: especially your viewers and the Geo community,
• 16:50
  SPEAKER_01: I will say, are familiar with the stack use of imagery.
• 16:55
  SPEAKER_01: But here we are implementing, we are adapting a stack extension, a stack MLM.
• 17:01
  SPEAKER_01: And we are describing all of those post-processing functions,
• 17:05
  SPEAKER_01: what are the Docker image that is required in order to run that model.
• 17:08
  SPEAKER_01: So each model is expected to run independently.
• 17:12
  SPEAKER_01: And since model weights should be also committed into something like hugging face
• 17:17
  SPEAKER_01: while I can have a hash of that model so it can be reproduced.
• 17:22
  SPEAKER_01: And in case something has a new version, the pipeline still runs and also distributed.
• 17:29
  SPEAKER_01: Same goes with it.
• 17:30
  SPEAKER_01: It should be described as what kind of training data does the model use.
• 17:36
  SPEAKER_01: What are the requirements and the restrictions for the model?
• 17:40
  SPEAKER_01: It should be described as a stack MLM item.
• 17:42
  SPEAKER_01: So it's in,
• 17:43
  SPEAKER_01: I will say,
• 17:44
  SPEAKER_01: it's understandable by everybody and library like .Geo natively supports stack MLM.
• 17:54
  SPEAKER_01: Yeah, we adapted this into our infrastructure,
• 17:58
  SPEAKER_01: and it is incorporated with GenML,
• 18:01
  SPEAKER_01: and we have Kubernetes that picks up those pipelines and the contract,
• 18:06
  SPEAKER_01: and then GenMLs kind of orchestrate the pipeline whenever people want to use
• 18:11
  SPEAKER_01: that model.
• 18:12
  SPEAKER_01: They discover it using the stack, and it also comes with, I think, Local version of it.
• 18:20
  SPEAKER_01: For example,
• 18:21
  SPEAKER_01: some models are meant to be worked only for a certain geographic reason,
• 18:25
  SPEAKER_01: and they would be described in a stack MLM that this is the geographic region
• 18:29
  SPEAKER_01: that this model was trained on and it is expected to work on this one.
• 18:33
  SPEAKER_01: So for people who are trying to find a model for that specific geographical region,
• 18:37
  SPEAKER_01: STAT will allow them to discover those kind of models and then they can run
• 18:42
  SPEAKER_01: and execute and try to see how it works on their area of interest,
• 18:46
  SPEAKER_01: on their imagery per se.
• 18:51
  SPEAKER_01: Maybe I'll explain it in a word,
• 18:53
  SPEAKER_01: but more of a concept how all of this comes together and how we pick it in
• 18:58
  SPEAKER_01: the infrastructure.
• 18:58
  SPEAKER_01: So each model is agnostic enough so it can run its own.
• 19:03
  SPEAKER_01: So we don't need to,
• 19:04
  SPEAKER_01: it doesn't break anything else and people can switch and discover the models.
• 19:09
  SPEAKER_00: Fantastic.
• 19:10
  SPEAKER_00: Yeah, as you mentioned,
• 19:11
  SPEAKER_00: Stack was created for imagery,
• 19:14
  SPEAKER_00: but it solved this problem of discovery and reuse and it's now being applied
• 19:18
  SPEAKER_00: to models, which I think is a fantastic concept.
• 19:20
  SPEAKER_00: And it's great to see it being adopted.
• 19:23
  SPEAKER_00: So the idea here is people produce a standardized pipeline and others can reuse it, extend it.
• 19:28
  SPEAKER_00: And it's also integrated into a platform.
• 19:32
  SPEAKER_01: Yes, that's correct.
• 19:33
  SPEAKER_01: There's a platform that we have been,
• 19:35
  SPEAKER_01: it's an initiative that HART has started, free and open source AI.
• 19:41
  SPEAKER_01: It's called FAIR.
• 19:42
  SPEAKER_01: It's a platform that is a first consumer of this stack item,
• 19:46
  SPEAKER_01: but definitely,
• 19:47
  SPEAKER_01: as you know, stack, I mean,
• 19:49
  SPEAKER_01: we can use other browsers as well,
• 19:50
  SPEAKER_01: but the current stack browsers do not have inference engine, per se, unless, like,
• 19:57
  SPEAKER_01: that's the next feature that we would like to work on is how to
• 20:00
  SPEAKER_01: When there are models that are encoded in stack MLM items,
• 20:03
  SPEAKER_01: maybe we can also run them as a browser, in a browser.
• 20:07
  SPEAKER_01: So that would be very nice to do.
• 20:08
  SPEAKER_01: But currently,
• 20:10
  SPEAKER_01: the platform that we have,
• 20:12
  SPEAKER_01: probably I can showcase later on how it works,
• 20:15
  SPEAKER_01: how we discover the models,
• 20:18
  SPEAKER_01: how it is encoded,
• 20:19
  SPEAKER_01: and people can go there and check out the models that are available already in
• 20:23
  SPEAKER_01: the stack collection and check it out and run it on their area.
• 20:28
  SPEAKER_00: Fantastic.
• 20:29
  SPEAKER_00: And is that a hosted platform?
• 20:31
  SPEAKER_00: Is that something that people run themselves?
• 20:35
  SPEAKER_01: Currently, I would say Boss.
• 20:37
  SPEAKER_01: It's an open source tool,
• 20:38
  SPEAKER_01: so you can host it on your own if you like and build your own stack collection
• 20:43
  SPEAKER_01: and maintain like XYZ different collections of the model.
• 20:48
  SPEAKER_01: With this current open call,
• 20:50
  SPEAKER_01: we are expecting We're using at least 20 different models to be incorporated
• 20:54
  SPEAKER_01: within this stack, so it will be available for everybody.
• 20:57
  SPEAKER_01: And there is also a hosted version of this that people can use,
• 21:02
  SPEAKER_01: which HART maintains, of course, for the humanitarian use cases.
• 21:08
  SPEAKER_01: There is a dev version and there is a production version.
• 21:14
  SPEAKER_01: The dev version is much more updated where we are currently working
• 21:18
  SPEAKER_01: on this Kubernetes environment so we can run any model.
• 21:23
  SPEAKER_01: Production version is going to be rolled out within this month with some of
• 21:28
  SPEAKER_01: the models that we already have.
• 21:30
  SPEAKER_01: By the end of the year,
• 21:31
  SPEAKER_01: when we have more models from the community,
• 21:34
  SPEAKER_01: Then it will be available in that hosted version as well.
• 21:37
  SPEAKER_00: Fantastic.
• 21:37
  SPEAKER_00: So just so people know the context, this is October 26th.
• 21:40
  SPEAKER_00: So we're talking basically by 2027,
• 21:42
  SPEAKER_00: this platform will be available for people to go in and see their models in use.
• 21:50
  SPEAKER_00: One other sort of requirement, I noticed you're looking for higher resolution imagery, really.
• 21:56
  SPEAKER_00: It appears to be one meter or better.
• 21:57
  SPEAKER_00: Is that right?
• 22:00
  SPEAKER_01: Yes, that's correct.
• 22:02
  SPEAKER_01: And currently,
• 22:04
  SPEAKER_01: because this also comes because of the area and nature of the work
• 22:09
  SPEAKER_01: that we have worked with in the past,
• 22:12
  SPEAKER_01: we had connections before with different Satellite imagery providers
• 22:17
  SPEAKER_01: who would usually donate that imagery when there is a disaster and it's
• 22:21
  SPEAKER_01: a high resolution.
• 22:22
  SPEAKER_01: And over the years,
• 22:23
  SPEAKER_01: like OpenStreetMap community and us,
• 22:26
  SPEAKER_01: we work at looking at the high resolution,
• 22:28
  SPEAKER_01: tracing the building footprints or tracing the map features.
• 22:31
  SPEAKER_01: And this is kind of,
• 22:33
  SPEAKER_01: I would say,
• 22:33
  SPEAKER_01: fall into the same domain and hence expanded to the AI work that we have done.
• 22:40
  SPEAKER_01: However,
• 22:40
  SPEAKER_01: the platform that we are building is not restricted to the type of imagery
• 22:47
  SPEAKER_01: by design, but we wanted to make it specific to let's first focus on the VHR.
• 22:53
  SPEAKER_01: I would say very high-resolution imagery,
• 22:55
  SPEAKER_01: and then later on,
• 22:56
  SPEAKER_01: like maybe next year,
• 22:57
  SPEAKER_01: we've expanded to support also Sentinel, Sentinel-2 imagery as well.
• 23:03
  SPEAKER_01: But currently, you are right that it is less than one meter of VHR RGB to be very specific.
• 23:11
  SPEAKER_00: Yes, that's it.
• 23:11
  SPEAKER_00: One-meter RGB,
• 23:13
  SPEAKER_00: probably satellite imagery to be released by one of the big providers after
• 23:17
  SPEAKER_00: an incident.
• 23:20
  SPEAKER_00: Fantastic.
• 23:20
  SPEAKER_00: We're coming up to time.
• 23:22
  SPEAKER_00: Maybe Petri would like to finish off just by explaining why people would like to engage.
• 23:27
  SPEAKER_00: And I also know that there is some funding available for people that want to contribute models.
• 23:34
  SPEAKER_02: Yeah, so I think call to action, ways to get involved.
• 23:38
  SPEAKER_02: I know probably the listeners come with different backgrounds.
• 23:41
  SPEAKER_02: So say if somebody is also not as technical,
• 23:44
  SPEAKER_02: we always welcome people who want to learn about mapping,
• 23:47
  SPEAKER_02: contribute to all the activations.
• 23:50
  SPEAKER_02: You can go on our website and go from there.
• 23:53
  SPEAKER_02: I think specifically for the developers,
• 23:56
  SPEAKER_02: data scientists, first of all, obviously the two codes we've mentioned.
• 24:01
  SPEAKER_02: One is closed already.
• 24:03
  SPEAKER_02: But we do have the one for GOAI models and one for like community,
• 24:08
  SPEAKER_02: for organizations that can run.
• 24:11
  SPEAKER_02: We can maybe share the links for those two calls that are still open and people can check those.
• 24:17
  SPEAKER_02: Those are the ones that have some funding available.
• 24:22
  SPEAKER_02: But other than that,
• 24:23
  SPEAKER_02: I think the one thing that I always say is not just for our AI-assisted service and
• 24:29
  SPEAKER_02: the platform for all the tools that we didn't cover today that allow this kind
• 24:33
  SPEAKER_02: of full end-to-end mapping journey.
• 24:37
  SPEAKER_02: They're all open source.
• 24:38
  SPEAKER_02: They're all on GitHub.
• 24:39
  SPEAKER_02: So for anyone who wants to go to GitHub at HotOSM and all the repos,
• 24:44
  SPEAKER_02: we also welcome contributors at any time.
• 24:48
  SPEAKER_02: They can, you know, look at some, we have a number of projects people can contribute.
• 24:53
  SPEAKER_02: And the last thing,
• 24:54
  SPEAKER_02: if you're interested,
• 24:56
  SPEAKER_02: we do working groups,
• 24:57
  SPEAKER_02: virtual sessions in two time zones that I run for tech and innovation.
• 25:01
  SPEAKER_02: Again, details on our website.
• 25:03
  SPEAKER_02: If you just want to connect and ask many more questions,
• 25:06
  SPEAKER_02: you're not sure where to start, working groups and Slack.
• 25:09
  SPEAKER_02: So various routes to get involved.
• 25:13
  SPEAKER_00: Okay.
• 25:13
  SPEAKER_00: Well, it's been such a pleasure to have you both on the podcast.
• 25:17
  SPEAKER_00: Today,
• 25:18
  SPEAKER_00: look forward to hearing the updates on the project and wish you all the best with
• 25:21
  SPEAKER_00: the calls.
• 25:23
  SPEAKER_02: Thank you, Robin.
• 25:24
  SPEAKER_02: Thank you so much for having us.
• 25:26
  SPEAKER_02: Yeah.
• 25:27
  SPEAKER_00: Bye, then.
• 25:29
  SPEAKER_02: Bye.
• 25:29
  SPEAKER_02: Bye, everyone.