Latest / Satellite image deep learning / Democratising access to GeoAI with InstaGeo

Transcript

Democratising access to GeoAI with InstaGeo · Satellite image deep learning

• 0:04
  SPEAKER_00: Hi, Ibrahim.
• 0:05
  SPEAKER_00: Welcome to the podcast.
• 0:07
  SPEAKER_00: Yeah.
• 0:08
  SPEAKER_00: Hi, Robin.
• 0:08
  SPEAKER_00: Yeah.
• 0:09
  SPEAKER_00: Thank you for having me.
• 0:10
  SPEAKER_00: Brilliant.
• 0:11
  SPEAKER_00: Let's begin with where you work and what you do.
• 0:15
  SPEAKER_01: Yeah.
• 0:17
  SPEAKER_01: Yeah.
• 0:17
  SPEAKER_01: Yeah.
• 0:18
  My name is Ibrahim and I'm a senior AI research engineer
• 0:24
  leading the team at Kigali.
• 0:28
  SPEAKER_01: Yeah.
• 0:29
  SPEAKER_01: I work for InstaDeep.
• 0:31
  And my team is focused mainly on the AI for social good
• 0:36
  projects that we do at Instagio.
• 0:38
  SPEAKER_01: So,
• 0:39
  SPEAKER_01: yeah,
• 0:40
  SPEAKER_01: the main thing that's brought us here today,
• 0:42
  SPEAKER_01: which is one of the projects that we've been working on,
• 0:44
  SPEAKER_01: which is Instagio.
• 0:47
  SPEAKER_00: Brilliant.
• 0:47
  SPEAKER_00: Why don't you tell us a little bit about that then?
• 0:51
  SPEAKER_01: Yeah, maybe just a couple of steps back to talk about history.
• 0:57
  SPEAKER_01: So in the year 2019 slash 2020,
• 1:01
  there was a major low-cost plague that affected some parts of
• 1:07
  the Horn of Africa
• 1:09
  SPEAKER_01: and some parts of Asia.
• 1:10
  SPEAKER_01: It was a serious outbreak.
• 1:13
  SPEAKER_01: And at that time, when we realized it in my...
• 1:17
  uh company we were like okay how can we leverage our expertise
• 1:22
  in ai to uh address
• 1:25
  this kind of challenges so that it doesn't really occur like in
• 1:30
  the future or if it
• 1:31
  does the impact is minimal so then we jumped uh onto it and uh
• 1:37
  started looking into
• 1:38
  SPEAKER_01: it reading a lot of
• 1:40
  SPEAKER_01: literature work around ecological modeling and low cost research.
• 1:45
  SPEAKER_01: And then, yeah, that was basically how we started.
• 1:48
  And because we wanted to build a tool that was going to be free
• 1:53
  for researchers as
• 1:56
  SPEAKER_01: well as practitioners to use,
• 1:59
  SPEAKER_01: we realized that from the onset,
• 2:01
  SPEAKER_01: we also needed to make sure that the data we consume,
• 2:05
  the tools that we use to consume the data and build the
• 2:10
  software is also free to
• 2:12
  SPEAKER_01: use.
• 2:12
  SPEAKER_01: And interestingly, there's a lot of
• 2:15
  SPEAKER_01: free and open source geospatial data,
• 2:19
  SPEAKER_01: especially multispectral data from the likes of Sentinel-2,
• 2:24
  SPEAKER_01: Landsat 8,
• 2:26
  SPEAKER_01: HLS,
• 2:27
  SPEAKER_01: and then the like of that.
• 2:29
  SPEAKER_01: And this kind of data, like I said, they are free and open to use.
• 2:33
  SPEAKER_01: They have tremendous use, I would say.
• 2:39
  So,
• 2:40
  but we realized something very critical,
• 2:44
  which was that this data is free to use.
• 2:47
  And a couple of researchers have also built what we call
• 2:52
  foundational models around this data.
• 2:54
  SPEAKER_01: So by foundational model,
• 2:55
  we mean they take the data and then they throw the very big
• 2:59
  models that land in an
• 3:01
  SPEAKER_01: unsupervised way,
• 3:02
  try to understand the data so that those models at the end of
• 3:05
  the day can be easily
• 3:06
  SPEAKER_01: fine tuned for different
• 3:08
  SPEAKER_01: downstream tasks.
• 3:10
  SPEAKER_01: Now,
• 3:11
  SPEAKER_01: what we found is that the data is there free to use,
• 3:14
  SPEAKER_01: the models,
• 3:17
  SPEAKER_01: foundational models are also there free to use,
• 3:20
  SPEAKER_01: but there is something missing in between,
• 3:23
  SPEAKER_01: which is that there is,
• 3:26
  SPEAKER_01: at least at the time when we realized this,
• 3:30
  SPEAKER_01: there is no free tool
• 3:32
  that allows you to take the ROS Sentinel-2 data or Landsat-8
• 3:38
  data and produce the
• 3:40
  same data that you will feed into just geospatial machine
• 3:43
  learning models.
• 3:44
  SPEAKER_01: And interestingly,
• 3:45
  all the researchers that also released their model code for the
• 3:49
  geospatial machine
• 3:50
  SPEAKER_01: learning model and say,
• 3:51
  SPEAKER_01: you can fine-tune it for various downstream tasks,
• 3:54
  SPEAKER_01: no one also releases
• 3:57
  the data pipeline code that was used to generate the data that
• 4:02
  was used to train the model.
• 4:04
  So what that now means is that we have this geospatial machine
• 4:07
  learning model.
• 4:08
  SPEAKER_01: We also have the free data,
• 4:12
  SPEAKER_01: but you can't really put it to use except you spend,
• 4:17
  SPEAKER_01: you invest a lot of time,
• 4:18
  SPEAKER_01: a lot of effort,
• 4:20
  and you have the technical expertise to be able to like move
• 4:23
  from having that free
• 4:25
  open source raw data to a model ready data that can be used for
• 4:30
  training and
• 4:31
  SPEAKER_01: inference.
• 4:33
  SPEAKER_01: This is sort of like how we started and then the challenge we realized.
• 4:37
  SPEAKER_00: Right.
• 4:37
  SPEAKER_00: So you were kind of under time pressure.
• 4:39
  SPEAKER_00: There's this locust outbreak.
• 4:40
  SPEAKER_00: You want to make a tool.
• 4:41
  Anybody can,
• 4:43
  you know,
• 4:43
  use models to find out where these locusts are affecting,
• 4:46
  etc.
• 4:47
  SPEAKER_00: You've seen that the data is out there.
• 4:49
  SPEAKER_00: The models are out there, but it's not in a usable format.
• 4:53
  SPEAKER_00: So you need to build some tooling to address that.
• 4:57
  SPEAKER_01: Yeah, although we actually realized this after a lot of work.
• 5:00
  So the first step that we did when we started looking into the
• 5:05
  locust breeding
• 5:05
  ground problem was to research like what are the techniques
• 5:09
  that researchers are
• 5:12
  SPEAKER_01: using to
• 5:13
  understand low-cost habitat and the breeding mechanism and then
• 5:19
  we look at we
• 5:20
  looked into the literature around what is called ecological
• 5:23
  niche modeling where we
• 5:25
  tried some classical machine learning algorithms and we tried
• 5:29
  some approaches to do
• 5:30
  SPEAKER_01: what is called absence data pseudo absence data sampling
• 5:36
  And then we moved from there to building a model that basically
• 5:44
  took variables,
• 5:47
  SPEAKER_01: different environmental and climatic variables like temperature,
• 5:51
  SPEAKER_01: pressure,
• 5:53
  SPEAKER_01: soil moisture content,
• 5:55
  SPEAKER_01: humidity level,
• 5:56
  SPEAKER_01: precipitation rate.
• 5:58
  SPEAKER_01: So we took all of those low level variables.
• 6:01
  SPEAKER_01: We put them together.
• 6:03
  SPEAKER_01: And then we fed it to our model.
• 6:04
  SPEAKER_01: So that was the second stage of the work that we did.
• 6:07
  SPEAKER_01: And we also published a paper on that.
• 6:12
  SPEAKER_01: And then when we finished the second stage,
• 6:14
  SPEAKER_01: then that was when we realized the first challenge that we had,
• 6:20
  SPEAKER_01: which was that
• 6:22
  SPEAKER_01: We built the model, but we cannot operationalize the model.
• 6:27
  SPEAKER_01: So by operationalization,
• 6:28
  I mean we cannot put it into everyday use because the data that
• 6:35
  we used was created
• 6:36
  SPEAKER_01: by aggregating different variables from different sources.
• 6:40
  SPEAKER_01: And those variables, they are not all available when you need them.
• 6:46
  SPEAKER_01: So for example,
• 6:47
  SPEAKER_01: if you need to make inference today,
• 6:48
  SPEAKER_01: you may realize that a couple of them are not available or if,
• 6:53
  SPEAKER_01: yeah,
• 6:53
  SPEAKER_01: it's very hard in any case,
• 6:55
  SPEAKER_01: like to assemble like all the different sources at the given time,
• 7:00
  SPEAKER_01: because some of them,
• 7:01
  SPEAKER_01: they release them like after some period of time.
• 7:05
  SPEAKER_01: So we started looking into,
• 7:06
  SPEAKER_01: okay,
• 7:07
  SPEAKER_01: this is a challenge and we don't want to just,
• 7:11
  SPEAKER_01: you know,
• 7:11
  SPEAKER_01: do research and publish paper,
• 7:12
  SPEAKER_01: but rather we want to have real impact on the ground.
• 7:16
  SPEAKER_01: So this was what informed our decision to look into multispectral data,
• 7:21
  SPEAKER_01: because we saw that multispectral data,
• 7:23
  SPEAKER_01: especially the HLS archive,
• 7:27
  SPEAKER_01: which was a harmonization between Sentinel-2 and Landsat-8,
• 7:30
  SPEAKER_01: is available every three to five days.
• 7:33
  SPEAKER_01: So that meant that we can easily get data anytime we need it.
• 7:39
  SPEAKER_01: So the question now at that time was,
• 7:45
  we've seen that multispectral data is useful for a lot of use
• 7:48
  cases like land cover
• 7:50
  SPEAKER_01: mapping,
• 7:52
  SPEAKER_01: crop classification and the likes.
• 7:54
  SPEAKER_01: But can we use it for identifying locust breeding grounds?
• 7:59
  SPEAKER_01: That was the first...
• 8:02
  SPEAKER_01: hypothesis that we came up with.
• 8:04
  SPEAKER_01: And interestingly,
• 8:05
  SPEAKER_01: when we tested that,
• 8:06
  SPEAKER_01: we found that,
• 8:07
  SPEAKER_01: yes,
• 8:07
  you can actually use multispectral data to identify local
• 8:12
  breeding grounds.
• 8:13
  SPEAKER_01: And that was a very great moment for us.
• 8:16
  SPEAKER_01: We actually published our finding.
• 8:19
  SPEAKER_01: Then after that,
• 8:21
  SPEAKER_01: that was when we now realized the initial thing I said,
• 8:25
  which was that there's a disconnect between the multispectral
• 8:28
  data and the
• 8:29
  SPEAKER_01: geospatial foundational models.
• 8:32
  So the effort it took us to build our first geospatial data
• 8:39
  pipeline that takes raw
• 8:42
  SPEAKER_01: multispectral data and convert it to model-ready format,
• 8:46
  SPEAKER_01: once we completed this,
• 8:47
  SPEAKER_01: I realized that,
• 8:48
  SPEAKER_01: look,
• 8:48
  SPEAKER_01: this is not a trivial thing,
• 8:49
  SPEAKER_01: and we are sure that a lot of people will not be able to do this.
• 8:54
  SPEAKER_01: Then we decided to say, okay, let's put things together
• 8:58
  and create a tool that allows researchers and practitioners to
• 9:03
  leverage this open
• 9:04
  SPEAKER_01: source multispectral data for any use case.
• 9:08
  SPEAKER_01: Because we've realized that this data has a lot of use cases.
• 9:11
  SPEAKER_01: But then these are some of the bottlenecks, like I mentioned.
• 9:15
  SPEAKER_01: So that was how we started building Instagio.
• 9:19
  SPEAKER_01: Yeah.
• 9:21
  SPEAKER_00: Good.
• 9:21
  SPEAKER_00: Can I just want to clarify,
• 9:23
  SPEAKER_00: was there a particular,
• 9:24
  was it foundational models in particular that you thought were
• 9:27
  challenging or just
• 9:28
  SPEAKER_00: any model and you decided foundational models were where you would focus?
• 9:35
  SPEAKER_01: You know,
• 9:35
  the good thing about the foundational models is that because
• 9:38
  they've already been
• 9:38
  SPEAKER_01: pre-trained on a lot of data,
• 9:40
  SPEAKER_01: so you can easily gear them for your downstream task of interest.
• 9:44
  That was the motivation to actually lean towards just special
• 9:49
  foundational models.
• 9:50
  SPEAKER_01: And also,
• 9:51
  one of the work that also inspired our decision on just special
• 9:55
  foundational models
• 9:55
  SPEAKER_01: is the work that...
• 9:57
  the collaboration between NASA and the IBM research team where
• 10:01
  they created a
• 10:03
  SPEAKER_01: family of geospatial financial models called Preetvy.
• 10:06
  SPEAKER_01: So they released version one of Preetvy.
• 10:08
  SPEAKER_01: They also released version two of Preetvy.
• 10:11
  SPEAKER_01: So like if you look at Instagio,
• 10:13
  SPEAKER_01: if you look at our repo,
• 10:15
  you will see that we've mostly relied on the Preetvy family of
• 10:19
  models to actually
• 10:20
  SPEAKER_01: derive our own models,
• 10:22
  SPEAKER_01: which we use on different downstream tasks.
• 10:26
  SPEAKER_00: Right.
• 10:26
  And they were one of the first sort of foundational models that
• 10:30
  were published.
• 10:33
  uh you mean the first foundational model was published in the
• 10:36
  space of uh well
• 10:37
  there was a lot of them but prithi got a lot of attention when
• 10:40
  it was announced yes
• 10:41
  and that also yes actually pre-trained on the hls i think so
• 10:45
  that's yes it was what
• 10:48
  you wanted right okay brilliant so you've identified the models
• 10:51
  identified the
• 10:51
  source uh how effective was the solution for the locust
• 10:55
  breeding ground detection
• 10:56
  SPEAKER_00: then
• 10:58
  SPEAKER_01: Yeah,
• 10:58
  so the first thing that we did was that we tried to replicate
• 11:02
  the results that we
• 11:03
  SPEAKER_01: got earlier with our new pipeline.
• 11:05
  SPEAKER_01: And yeah, we got a very interesting result.
• 11:09
  SPEAKER_01: Like we were able to replicate it and even improve on the result.
• 11:13
  SPEAKER_01: And currently at the moment, we are now looking into like having
• 11:18
  on the ground impact because you know the first thing for this
• 11:21
  kind of task is for
• 11:23
  you to build the model then you assess it on the ground and
• 11:27
  that's actually the
• 11:27
  very difficult part so for this we are actually talking to some
• 11:33
  partners who are
• 11:34
  SPEAKER_01: already like involved in the field work to see how we can collaborate to
• 11:40
  actually assess it on the ground because there's a limit to
• 11:44
  what we can do as researchers.
• 11:45
  SPEAKER_01: So we need to also like collaborate with other players.
• 11:50
  SPEAKER_00: Of course.
• 11:50
  SPEAKER_00: And what were some of the particular challenges there?
• 11:56
  SPEAKER_01: So early on,
• 11:57
  SPEAKER_01: we tried when we started looking into the problem,
• 12:00
  SPEAKER_01: we tried to do the research and then at the same time,
• 12:05
  SPEAKER_01: like get the model to be used by people in the field.
• 12:10
  SPEAKER_01: But it turns out that the logistical,
• 12:16
  SPEAKER_01: there's a lot of logistical challenge around this,
• 12:19
  SPEAKER_01: like collaborating with people like different organizations.
• 12:25
  SPEAKER_01: You need to first find, you know, a common interest.
• 12:29
  SPEAKER_01: And then from there, like, so it's been difficult.
• 12:32
  SPEAKER_01: And it's one thing that you've actually realized that it sounds trivial,
• 12:36
  SPEAKER_01: but at the same time,
• 12:37
  SPEAKER_01: it's not easy.
• 12:39
  SPEAKER_00: Right,
• 12:40
  SPEAKER_00: so these are organizations that had individuals going out,
• 12:43
  SPEAKER_00: making maps of,
• 12:45
  SPEAKER_00: you know,
• 12:45
  SPEAKER_00: where these breeding camps are occurring,
• 12:47
  SPEAKER_00: capturing the data,
• 12:49
  SPEAKER_01: sharing that with you somehow.
• 12:52
  SPEAKER_01: Yeah,
• 12:52
  SPEAKER_01: I must stress that when we started the work on Locust,
• 12:55
  SPEAKER_01: we actually,
• 12:56
  SPEAKER_01: we looked into the organizations to partner with,
• 13:00
  SPEAKER_01: and then we saw that the UN...
• 13:02
  Food and Agriculture Organization has been doing a lot of work
• 13:05
  in this regard over
• 13:07
  SPEAKER_01: a lot of years ago.
• 13:08
  SPEAKER_01: So they were a natural first partner.
• 13:11
  SPEAKER_01: So we reached out to them and then we started partnering with them.
• 13:15
  We talked to like a lot of professionals from FAO and we got a
• 13:21
  lot of insight.
• 13:22
  SPEAKER_01: We got a lot of
• 13:25
  uh yeah insights and guide on how to use the data because the
• 13:29
  ground truth data
• 13:30
  that we use comes from the fao archive that yeah it's called
• 13:33
  the locust uh uh
• 13:36
  locusts i forgot the name but locust something is the name of
• 13:41
  the archive that
• 13:41
  contains uh the data and also for the underground assessment uh
• 13:46
  they are also our
• 13:47
  SPEAKER_01: first partner to actually uh talk to
• 13:50
  SPEAKER_01: And yeah,
• 13:52
  SPEAKER_01: the talks is already ongoing because they have access also to,
• 13:56
  SPEAKER_01: you know,
• 13:56
  SPEAKER_01: the different countries where this low-cost plague usually occur.
• 14:03
  SPEAKER_01: And then they already have...
• 14:05
  a kind of system or let me say a networking place of different
• 14:09
  players within the
• 14:11
  field but just to mention that low cost was actually our first
• 14:16
  pioneer project but
• 14:18
  we've since expanded to even other other projects so like
• 14:24
  damage assessment
• 14:27
  SPEAKER_01: air quality assessment,
• 14:31
  SPEAKER_01: which is also called aerosol optical depth estimation,
• 14:34
  SPEAKER_01: then biomass estimation as well.
• 14:36
  SPEAKER_01: Because the good thing,
• 14:37
  like I said earlier about multispectral data is that it has a
• 14:40
  lot of use cases.
• 14:41
  SPEAKER_01: So we built one tool that can be used to solve any problem
• 14:47
  SPEAKER_01: that is possible to be solved using this multispectral data.
• 14:51
  SPEAKER_00: Right.
• 14:51
  SPEAKER_00: So that's what INSIDEEP is.
• 14:52
  It is the tooling that you need to address all these different
• 14:56
  challenges.
• 14:59
  SPEAKER_00: And so the model is trained.
• 15:02
  SPEAKER_00: You're happy with its performance.
• 15:04
  How have you been deploying that and putting that in the hands
• 15:06
  of the communities
• 15:07
  SPEAKER_00: that need that information?
• 15:10
  SPEAKER_01: Yeah,
• 15:11
  SPEAKER_01: before you go into that,
• 15:12
  SPEAKER_01: let me first give a high level picture of what the framework looks like.
• 15:16
  So Instagio is an end-to-end geospatial machine learning
• 15:19
  framework that makes it
• 15:20
  SPEAKER_01: very easy for researchers,
• 15:22
  SPEAKER_01: practitioners to go from raw satellite,
• 15:24
  SPEAKER_01: raw multispectral data to an operational model.
• 15:28
  SPEAKER_01: That's the goal.
• 15:30
  SPEAKER_01: And it has three core components.
• 15:32
  SPEAKER_01: The first component is the data components.
• 15:34
  SPEAKER_01: So this does the heavy lifting of taking the raw,
• 15:37
  SPEAKER_01: basically not only taking the raw satellite data,
• 15:39
  SPEAKER_01: all you feed to the data component is geolocation.
• 15:43
  SPEAKER_01: So a pair of longitude and latitude, date and label information.
• 15:46
  SPEAKER_01: And that's all it needs.
• 15:48
  SPEAKER_01: You can also feed in a raster.
• 15:50
  And then once you give it that and you configure the different
• 15:55
  parameters of the
• 15:55
  SPEAKER_01: pipeline,
• 15:56
  it will go to the HLS archive or Sentinel-2 archive or
• 15:59
  Sentinel-1 archive,
• 16:01
  SPEAKER_01: depending on the one you choose.
• 16:03
  It will fetch all the corresponding granules,
• 16:08
  put them together if you have a temporal data
• 16:12
  SPEAKER_01: and then crop the region that is necessary.
• 16:16
  And then at the end of the day,
• 16:18
  it gives you what we call chips in the shape that you
• 16:23
  specified.
• 16:23
  SPEAKER_01: So for example, if I requested for six bonds,
• 16:28
  SPEAKER_01: with three time steps and special size of 224 by 224.
• 16:31
  At the end of the day,
• 16:33
  it will give me a TIF file that has a shape of 3 by 6 by 224 by
• 16:41
  224.
• 16:42
  SPEAKER_01: So this data in this format is now ready for,
• 16:46
  SPEAKER_01: and also it's paired with the label,
• 16:48
  SPEAKER_01: which in our case is a segmentation map of size 224 by 224.
• 16:53
  SPEAKER_01: And then you can then feed this to a model.
• 16:56
  SPEAKER_01: So this goes to the next component, which is the model component.
• 16:59
  And this component serves to basically fine-tune your special
• 17:03
  machine learning
• 17:03
  SPEAKER_01: model or train one from scratch.
• 17:07
  And at the same time,
• 17:08
  you can also do inference after you have a trained model.
• 17:10
  SPEAKER_01: So then you can take the data from the data component,
• 17:14
  SPEAKER_01: use it in the model component,
• 17:15
  SPEAKER_01: and then you have a model checkpoint.
• 17:19
  SPEAKER_01: Then we have the other component, which is the application component.
• 17:24
  SPEAKER_01: Now, in this application component, you can take your...
• 17:29
  SPEAKER_01: inference, your predicted files basically.
• 17:34
  SPEAKER_01: So because after you perform inference,
• 17:36
  SPEAKER_01: you save a GOT file and then you can visualize it in the app component.
• 17:43
  So basically it will overlay everything on an interactive map
• 17:46
  and then you're able
• 17:47
  SPEAKER_01: to visualize it.
• 17:48
  So this was our version zero,
• 17:52
  I would say,
• 17:52
  the first version that we released.
• 17:57
  SPEAKER_01: in 2024, but since then we've made a lot of improvement.
• 18:01
  SPEAKER_01: So like two weeks ago,
• 18:04
  SPEAKER_01: we released a new version whereby in the data component now,
• 18:08
  SPEAKER_01: because formally in the data component,
• 18:10
  SPEAKER_01: we need,
• 18:11
  SPEAKER_01: we usually pull the entire granule,
• 18:13
  SPEAKER_01: which is very big.
• 18:14
  SPEAKER_01: It has special size of 3660 by 3660 pixel.
• 18:17
  SPEAKER_01: We pull this and then we do like all the processing.
• 18:24
  SPEAKER_01: So for example,
• 18:25
  SPEAKER_01: we realize if you want to create a dataset that is just,
• 18:29
  SPEAKER_01: that will end up being 225 gigabyte,
• 18:32
  SPEAKER_01: for example,
• 18:33
  SPEAKER_01: you may end up pulling raw satellite data of like one terabyte.
• 18:36
  SPEAKER_01: So this was not scalable for us.
• 18:39
  SPEAKER_01: So now in the new version, what we've done is to leverage the stack API.
• 18:44
  which allows us to basically use cloud-optimized GOTF to only
• 18:50
  fetch the region of
• 18:51
  SPEAKER_01: interest from each of those granules.
• 18:54
  SPEAKER_01: And now our pipeline is very fast.
• 18:57
  You can ask it to create all the chips covering the entire
• 19:05
  Africa continent,
• 19:05
  SPEAKER_01: for example,
• 19:07
  and then you give that to it and just wait for it to give you
• 19:10
  the results.
• 19:11
  SPEAKER_01: So it's very easy to use and very fast.
• 19:13
  and it's quite modular as well because you could change your
• 19:17
  stack catalog use a
• 19:18
  different one but still use the processing you could take your
• 19:22
  tiffs and put them
• 19:23
  somewhere exactly exactly you got it you got it you got it then
• 19:27
  in the model
• 19:27
  SPEAKER_01: component as well we've tried to make our models very lightweight
• 19:32
  So like we do compression and then we track the carbon emission
• 19:37
  of our model as
• 19:40
  SPEAKER_01: well as the computational complexity.
• 19:43
  SPEAKER_01: So we have what we call a technique,
• 19:45
  SPEAKER_01: we call task specific distillation,
• 19:48
  whereby we create a final model that corresponds to the
• 19:54
  difficulty of the task.
• 19:56
  SPEAKER_01: So this helps us to have lightweight models.
• 19:59
  Because remember that the goal is for us to be able to
• 20:02
  operationalize the model so
• 20:03
  SPEAKER_01: that people can do continental scale inference.
• 20:06
  So we need something that is lightweight at the same time with
• 20:10
  good performance.
• 20:12
  SPEAKER_01: Then in the first version of the app component,
• 20:16
  SPEAKER_01: you needed to first do inference with the model components,
• 20:20
  SPEAKER_01: then you save the prediction geotiffs.
• 20:22
  Then you can then load them in the apps component and visualize
• 20:25
  it on the map.
• 20:27
  SPEAKER_01: But now in the new version of our apps component,
• 20:29
  SPEAKER_01: it's actually a full stack application built with the React framework.
• 20:34
  SPEAKER_01: And then it combines or it joins together the data component,
• 20:40
  SPEAKER_01: the model component,
• 20:41
  SPEAKER_01: and then the app component itself.
• 20:43
  So now if you go to Instagio.io,
• 20:45
  you are able to draw an area of interest on the map.
• 20:55
  SPEAKER_01: select some few parameters for the pipeline, for the data pipeline.
• 21:01
  SPEAKER_01: Then you click on the run job.
• 21:04
  SPEAKER_01: And okay, you also select like the model.
• 21:06
  SPEAKER_01: We have a couple of models in our model registry.
• 21:10
  So you select the model of interest after selecting your area
• 21:13
  of interest,
• 21:13
  SPEAKER_01: and then you click on run job.
• 21:15
  SPEAKER_01: And then it will go through the entire process.
• 21:17
  SPEAKER_01: So it will fetch the data sets, create the data in a model ready format.
• 21:22
  do inference and visualize the result on the map and also give
• 21:27
  you some analytics
• 21:28
  SPEAKER_01: that you can download and save as a PDF file.
• 21:31
  SPEAKER_01: So that's the current state of what we have.
• 21:34
  SPEAKER_00: Well, congratulations.
• 21:35
  It sounds very fully featured and it's obviously been the
• 21:40
  outcome of a lot of awful lot of work.
• 21:42
  SPEAKER_00: I'm wondering what you still feel is left to do.
• 21:47
  Yeah,
• 21:48
  I think for us,
• 21:51
  what is left to do is to have the actual impact on the ground.
• 21:55
  SPEAKER_01: And like I said, this is an entirely different process.
• 22:02
  We need to reach out to partners and see how we can leverage
• 22:08
  this open source tool
• 22:10
  that we've created to really have impact on the ground across
• 22:13
  different ways
• 22:15
  SPEAKER_01: sectors, I'll say.
• 22:18
  SPEAKER_00: Right.
• 22:18
  SPEAKER_00: So you want to see it out there being used as widely as possible.
• 22:24
  SPEAKER_00: Very exciting stuff.
• 22:26
  SPEAKER_00: The code itself is on GitHub.
• 22:28
  SPEAKER_00: That's open for everybody.
• 22:29
  We're going to do a separate video to have a look at Instagio
• 22:33
  and just walk through
• 22:34
  SPEAKER_00: some of those features.
• 22:36
  SPEAKER_00: I'll wrap this one up there just finally.
• 22:38
  SPEAKER_00: So obviously you can go to GitHub and follow the activities.
• 22:41
  SPEAKER_00: Do you post updates on LinkedIn or elsewhere?
• 22:45
  Yes,
• 22:45
  and on the website also,
• 22:47
  we have a feedback form for people to leave us feedback.
• 22:51
  SPEAKER_01: And you can also create issues on GitHub.
• 22:54
  SPEAKER_01: And yeah, we'll be ready to take a look at any issue that anyone has.
• 22:58
  SPEAKER_00: Brilliant.
• 22:58
  SPEAKER_00: Okay, well, I'll leave all the links in the show notes.
• 23:00
  SPEAKER_00: Once again, it's been a pleasure.
• 23:02
  SPEAKER_00: Until next time, take care.
• 23:04
  SPEAKER_00: Thank you.
• 23:04
  SPEAKER_00: Bye.