Latest / Satellite image deep learning / The FLAIR land cover mapping challenge

Transcript

The FLAIR land cover mapping challenge · Satellite image deep learning

• 0:05
  SPEAKER_00: Hello, I'm Robin Cole, and welcome to the Satellite Image Deep Learning Podcast.
• 0:09
  SPEAKER_00: In this episode,
• 0:10
  SPEAKER_00: I caught up with Nicholas Gonthier to learn about the FLARE Land Cover Mapping Challenge.
• 0:14
  SPEAKER_00: In this challenge, 20 centimeter resolution drawing imagery was used to create high quality annotations.
• 0:20
  SPEAKER_00: This data was paired with a time series of medium resolution Sentinel-2 images to
• 0:25
  SPEAKER_00: create a rich multidimensional data set.
• 0:28
  SPEAKER_00: Participants in the challenge were able to surpass the baseline solution by 10
• 0:32
  SPEAKER_00: points in the target metric,
• 0:33
  SPEAKER_00: representing a significant step forward in land cover classification capabilities.
• 0:38
  SPEAKER_00: The dataset is now being expanded to cover a larger area and incorporate additional
• 0:42
  SPEAKER_00: imaging modalities,
• 0:44
  SPEAKER_00: which have been shown to improve performance on this task.
• 0:47
  SPEAKER_00: Nicholas also provided important context about the objectives of the organization
• 0:51
  SPEAKER_00: running this challenge,
• 0:52
  SPEAKER_00: such as the need to balance model performance with processing costs.
• 0:56
  SPEAKER_00: This was a fascinating conversation, and I hope you enjoy this episode.
• 1:00
  SPEAKER_00: If you are watching on YouTube,
• 1:01
  SPEAKER_00: please like and subscribe,
• 1:02
  SPEAKER_00: as this will really help to grow the channel and reach a wider audience.
• 1:05
  SPEAKER_00: Thank you.
• 1:06
  SPEAKER_00: Hi, Nicholas.
• 1:09
  SPEAKER_00: Welcome to the podcast.
• 1:10
  SPEAKER_01: Hi, Robin.
• 1:11
  SPEAKER_01: Thank you for welcoming me.
• 1:13
  SPEAKER_00: It's a pleasure.
• 1:13
  SPEAKER_00: Let's begin with where you work and what you do.
• 1:16
  SPEAKER_01: Okay, so I'm a project manager in the innovation team at IGEN.
• 1:22
  SPEAKER_01: So IGEN is a national agency,
• 1:25
  SPEAKER_01: nonprofit one in France,
• 1:27
  SPEAKER_01: working about geographical and forestry information.
• 1:31
  SPEAKER_01: And so we are creating a lot of data for different policy,
• 1:36
  SPEAKER_01: public policy and policymakers,
• 1:38
  SPEAKER_01: mainly about agricultural information,
• 1:41
  SPEAKER_01: forestry,
• 1:42
  SPEAKER_01: ecological and land use.
• 1:44
  SPEAKER_00: Mm-hmm.
• 1:46
  SPEAKER_01: I'm more working about deep learning for observation and transduction.
• 1:51
  SPEAKER_00: Fantastic.
• 1:53
  SPEAKER_00: That's a nice introduction to the background.
• 1:56
  SPEAKER_00: And you've been involved in the Flare Challenge.
• 1:58
  SPEAKER_00: Do you mind giving us the background to that?
• 2:01
  SPEAKER_01: Yeah,
• 2:02
  SPEAKER_01: so in fact,
• 2:03
  SPEAKER_01: we create these very high quality data sets that we've called Flare about land
• 2:09
  SPEAKER_01: cover semantic segmentation in order to help us at Hygen to produce land cover projects,
• 2:16
  SPEAKER_01: especially products that can be deployed at country scale.
• 2:21
  SPEAKER_01: So it's now been used in the production pipeline at the Institute.
• 2:27
  SPEAKER_01: in order to have a very precise land cover.
• 2:29
  SPEAKER_01: And for that,
• 2:30
  SPEAKER_01: we have annotated a lot of different images from different regions of France in
• 2:34
  SPEAKER_01: order to have broad diversity.
• 2:36
  SPEAKER_01: And we then trained a supervised model on this data in order to help us to produce
• 2:45
  SPEAKER_01: this kind of mapping of the territory.
• 2:48
  SPEAKER_01: And we then decided to open both the datasets
• 2:53
  SPEAKER_01: and the code and the model in order to be able to reproduce.
• 2:58
  SPEAKER_01: And finally,
• 2:58
  SPEAKER_01: we organized two different challenges in order to encourage community,
• 3:05
  SPEAKER_01: more really computer vision and remote sensing community to work on our problem.
• 3:11
  SPEAKER_00: Fantastic.
• 3:12
  SPEAKER_00: So I think the first version of the competition and the dataset was 2022,
• 3:15
  SPEAKER_00: and then there was another version a year or two later.
• 3:18
  SPEAKER_00: What changed between those two versions?
• 3:21
  SPEAKER_01: Between the two versions,
• 3:23
  SPEAKER_01: we changed,
• 3:24
  SPEAKER_01: we added a new modality,
• 3:26
  SPEAKER_01: what we call a new modality in the second part of the dataset.
• 3:30
  SPEAKER_01: So the first dataset is composed of very high resolution images, parallel images.
• 3:36
  SPEAKER_01: that have been acquired by IGN.
• 3:40
  SPEAKER_01: So it's a 20 pixel ground level pixel with annotation made by hand and that have been verified.
• 3:48
  SPEAKER_01: So it's very high quality that we need to try and supervise model.
• 3:52
  SPEAKER_01: And on the second dataset,
• 3:53
  SPEAKER_01: we also add for each of the patch from these images,
• 3:57
  SPEAKER_01: Sentinel-2 time series from the same zone,
• 4:02
  SPEAKER_01: in order to bring more information about,
• 4:05
  SPEAKER_01: for instance,
• 4:05
  SPEAKER_01: phenology of vegetation,
• 4:07
  SPEAKER_01: or to be able to distinguish
• 4:09
  SPEAKER_01: some different classes that you can have between,
• 4:11
  SPEAKER_01: for example,
• 4:11
  SPEAKER_01: bare soil and agricultural,
• 4:13
  SPEAKER_01: because the behavior will be the same during the year.
• 4:17
  SPEAKER_01: And in fact,
• 4:17
  SPEAKER_01: the zones we provide are even wider than the small hourly zone in order to also
• 4:24
  SPEAKER_01: have context information.
• 4:25
  SPEAKER_01: Maybe you can know that if around you have a can of sea,
• 4:28
  SPEAKER_01: you can think that this is more sane than something else.
• 4:33
  SPEAKER_00: Right, so a high-resolution snapshot and then a time series at low resolution to...
• 4:38
  SPEAKER_00: Understand temporal signatures.
• 4:41
  SPEAKER_00: Exactly.
• 4:42
  SPEAKER_00: They're interesting.
• 4:42
  SPEAKER_00: And how do you exploit both of those image sequences in a model?
• 4:50
  SPEAKER_01: So we propose the baseline in this case to do this kind of fusion of information.
• 4:56
  SPEAKER_01: So we have first kind of classical units,
• 5:00
  SPEAKER_01: segmentation model for dealing with the high resolution IR images.
• 5:05
  SPEAKER_01: And then we use what's called a UTA.
• 5:08
  SPEAKER_01: So it's kind of units that are modified in order to have a temporal attention
• 5:13
  SPEAKER_01: system in order to compress and in the clever way,
• 5:17
  SPEAKER_01: the temporal information from the time series.
• 5:20
  SPEAKER_01: And we inject this information at different points in the main branch of the unit.
• 5:26
  SPEAKER_01: And we supervised the both branches,
• 5:28
  SPEAKER_01: so the one with the Arial part and the time series part,
• 5:33
  SPEAKER_01: with the high resolution annotation we have provided.
• 5:38
  SPEAKER_01: So that's the baseline we proposed.
• 5:40
  SPEAKER_01: And in the second challenge,
• 5:41
  SPEAKER_01: people try to do different kind of fusion elements,
• 5:45
  SPEAKER_01: try to remove some,
• 5:47
  SPEAKER_01: for instance,
• 5:48
  SPEAKER_01: cloudy images or try to only use very few images and also try to calibrate the
• 5:54
  SPEAKER_01: images based on the time series and everything.
• 5:57
  SPEAKER_01: A lot of things have been tried.
• 6:00
  SPEAKER_00: Fantastic.
• 6:01
  SPEAKER_00: And how do people do with their approaches versus the baseline?
• 6:08
  SPEAKER_01: The people bring,
• 6:09
  SPEAKER_01: I think if I remember well,
• 6:10
  SPEAKER_01: it was like something like almost 10 points in mean average precision in total.
• 6:16
  SPEAKER_01: So it was interesting things, and we will try to use them as much as possible.
• 6:22
  SPEAKER_01: In the second challenge,
• 6:23
  SPEAKER_01: we also add an extra thing,
• 6:25
  SPEAKER_01: the fact that we require user,
• 6:29
  SPEAKER_01: sorry,
• 6:29
  SPEAKER_01: participant,
• 6:30
  SPEAKER_01: to not use too much computer program.
• 6:33
  SPEAKER_01: Because, in fact, in the first challenge, we realized that some of the best methods were very costly.
• 6:39
  SPEAKER_01: And as we need to be able to provide data at a country scale,
• 6:45
  SPEAKER_01: so like a very huge amount of data,
• 6:49
  SPEAKER_01: we need to have efficient solutions.
• 6:52
  SPEAKER_01: So that's also one of the things we had in the contest.
• 6:55
  SPEAKER_00: Interesting.
• 6:55
  SPEAKER_00: And how did you specify the limitation there, was it?
• 6:58
  SPEAKER_01: We specified the limitation like,
• 7:01
  SPEAKER_01: on one review to have something like no more than twice the inference time from the
• 7:07
  SPEAKER_01: baseline model.
• 7:09
  SPEAKER_00: Right, so okay.
• 7:10
  SPEAKER_00: You set the hardware and said this is the expectations on how quickly it should run.
• 7:14
  SPEAKER_00: Okay, very interesting.
• 7:16
  SPEAKER_00: Okay.
• 7:16
  SPEAKER_00: I'm curious,
• 7:17
  SPEAKER_00: did you do any kind of investigation to see what was the most significant features
• 7:21
  SPEAKER_00: for particular classes?
• 7:22
  SPEAKER_00: Because you could imagine something like urban areas wouldn't actually matter to
• 7:26
  SPEAKER_00: have the time series,
• 7:27
  SPEAKER_00: but other like vegetation might.
• 7:30
  SPEAKER_01: So we remember that the vegetation learned a lot about the time series, but also the urban area.
• 7:38
  SPEAKER_01: We were surprised about that.
• 7:39
  SPEAKER_01: For instance, building by using Sentinel-2 time series get three points EOE improvement.
• 7:46
  SPEAKER_01: So the classes that have been the most improved are more the bare soil class,
• 7:51
  SPEAKER_01: in fact,
• 7:52
  SPEAKER_01: which should be more easily distinguished from agricultural land,
• 7:55
  SPEAKER_01: for instance.
• 7:57
  SPEAKER_01: And also vineyards have been also quite improved by this multimodal models.
• 8:03
  SPEAKER_00: Very interesting.
• 8:03
  SPEAKER_00: So there's some confused classes and having the time series actually helps to separate those two.
• 8:09
  SPEAKER_00: Okay, yeah, that makes sense.
• 8:12
  SPEAKER_00: And I understand there's also a third stage in the dataset and competition?
• 8:19
  SPEAKER_01: So, not in the competition, but in fact, we, so, this, providing this, no, sorry, I'll start again.
• 8:29
  SPEAKER_01: Organizing the competition forces us to have very clear baseline,
• 8:33
  SPEAKER_01: to have something properly visible,
• 8:35
  SPEAKER_01: and to be able to open both the dataset,
• 8:38
  SPEAKER_01: but also the code to be able to run and train model.
• 8:43
  SPEAKER_01: And then we also decided to publish in an AI conference in NeurIPS last year this
• 8:49
  SPEAKER_01: dataset in order to encourage academic people to work on our problems,
• 8:55
  SPEAKER_01: on our issue we have at IGN.
• 8:57
  SPEAKER_01: And for now, we are trying to think about creating a new version of the data sets.
• 9:03
  SPEAKER_01: Yes, we have three times more images with very high resolution.
• 9:09
  SPEAKER_01: So for now,
• 9:09
  SPEAKER_01: we have something like 800 kilometers squares annotated,
• 9:15
  SPEAKER_01: and we are aiming at 2,500 kilometers.
• 9:20
  SPEAKER_01: annotating and we will also add more modality or extra data so like sentinel 2
• 9:28
  SPEAKER_01: information but also sentinel 1 data on the same area right so i understand
• 9:34
  SPEAKER_00: expanding the modalities the number of sites presumably yes and the intuition is
• 9:39
  SPEAKER_00: that just with a larger data set you'll get better better models and better
• 9:43
  SPEAKER_01: performance exactly yeah what would you expect
• 9:46
  SPEAKER_01: And also,
• 9:47
  SPEAKER_01: and so,
• 9:47
  SPEAKER_01: yeah,
• 9:48
  SPEAKER_01: and we also released,
• 9:50
  SPEAKER_01: after the challenge,
• 9:51
  SPEAKER_01: we also released the baseline model with the weight so people can use them in order
• 9:55
  SPEAKER_01: to do their own things.
• 9:59
  SPEAKER_01: So they can fine tune them,
• 10:00
  SPEAKER_01: they can use them as evaluation,
• 10:02
  SPEAKER_01: as our initialization for training or some different things.
• 10:08
  SPEAKER_01: And we will do the same with the next version of the dataset for sure.
• 10:13
  SPEAKER_00: Fantastic.
• 10:14
  SPEAKER_00: I'd like to just talk a little bit about the process of creating these data sets.
• 10:17
  SPEAKER_00: Do you mind just at a high level saying how you go about it,
• 10:20
  SPEAKER_00: how you select where to sample data,
• 10:22
  SPEAKER_00: that kind of thing?
• 10:25
  SPEAKER_01: So, the idea was to have very diverse data sets, so we tried to sample data.
• 10:29
  SPEAKER_01: So,
• 10:30
  SPEAKER_01: it's only on the metropolitan areas,
• 10:34
  SPEAKER_01: so in parts of France and Europe,
• 10:36
  SPEAKER_01: because France also has a lot of different highlands in the Caribbean,
• 10:42
  SPEAKER_01: for instance,
• 10:42
  SPEAKER_01: or in the Indian Ocean.
• 10:44
  SPEAKER_01: And we will work on that later, for sure.
• 10:47
  SPEAKER_01: uh and uh then what we so we select a zone with different uh landscape with big
• 10:53
  SPEAKER_01: cities small cities uh with mountains and things like that and we also iterate on
• 10:58
  SPEAKER_01: that and for instance the next version of the set will have more diverse uh which
• 11:03
  SPEAKER_01: one again and we also this and then we uh make people annotating the different
• 11:10
  SPEAKER_01: classes the lan
• 11:11
  SPEAKER_01: covers after class.
• 11:12
  SPEAKER_01: And we also have a validation system that means that then a second person,
• 11:19
  SPEAKER_01: sometimes not always,
• 11:19
  SPEAKER_01: but sometimes at random,
• 11:21
  SPEAKER_01: will check if annotations are correct or not.
• 11:24
  SPEAKER_01: And if this is not the case,
• 11:26
  SPEAKER_01: we will go again on the annotation,
• 11:28
  SPEAKER_01: not to improve the quality of it,
• 11:29
  SPEAKER_01: because we aim quality of 95% accuracy on the human annotation.
• 11:40
  SPEAKER_01: It was very time-consuming and,
• 11:44
  SPEAKER_01: in fact,
• 11:44
  SPEAKER_01: costly,
• 11:45
  SPEAKER_01: but it's worth it because we were able to get a good AI model then.
• 11:51
  SPEAKER_00: Two questions.
• 11:52
  SPEAKER_00: On the selection,
• 11:54
  SPEAKER_00: did you go for an equal weighting on all the different classes,
• 11:57
  SPEAKER_00: or did you say,
• 11:57
  SPEAKER_00: this class is really hard,
• 12:00
  SPEAKER_00: or this location is hard,
• 12:00
  SPEAKER_00: let's have more data for that?
• 12:02
  SPEAKER_00: What's the strategy there?
• 12:04
  SPEAKER_01: At the beginning, we didn't do it like that.
• 12:06
  SPEAKER_01: We were more looking at the different regions of France because,
• 12:11
  SPEAKER_01: as we are a kind of mapping agency,
• 12:14
  SPEAKER_01: we already know where are the different,
• 12:16
  SPEAKER_01: let's say,
• 12:17
  SPEAKER_01: bioclimatic regions in France,
• 12:19
  SPEAKER_01: where we can find small cities,
• 12:21
  SPEAKER_01: more dense urban areas.
• 12:24
  SPEAKER_01: And then,
• 12:26
  SPEAKER_01: on the next version of the dataset,
• 12:27
  SPEAKER_01: we will focus specifically on some classes that are harder for the model,
• 12:31
  SPEAKER_01: yes.
• 12:33
  SPEAKER_01: We are able to say.
• 12:34
  SPEAKER_01: And for instance,
• 12:35
  SPEAKER_01: on the second challenge from Flare,
• 12:38
  SPEAKER_01: we decided to emphasize more about the validation because it was one of the hardest
• 12:43
  SPEAKER_01: things to distinguish.
• 12:44
  SPEAKER_01: And for instance,
• 12:45
  SPEAKER_01: we decided to put more validation class in the test set in order to select model
• 12:50
  SPEAKER_01: that will be able on this class.
• 12:51
  SPEAKER_00: Okay.
• 12:52
  SPEAKER_00: Yeah, it makes sense.
• 12:53
  SPEAKER_00: And in terms of determining which classes are hard,
• 12:55
  SPEAKER_00: did you see that it was always the same for the human and the model,
• 12:58
  SPEAKER_00: which classes were the hardest?
• 13:00
  SPEAKER_00: Or there are some classes that are easy for a model because it has a time series,
• 13:03
  SPEAKER_00: but hard for humans,
• 13:04
  SPEAKER_00: eh?
• 13:09
  SPEAKER_01: Try to think about it.
• 13:10
  SPEAKER_01: I think one of the classes that is...
• 13:11
  SPEAKER_01: So some of the hardest classes for machines are the classes like Brushwood and Plumland,
• 13:20
  SPEAKER_01: which are maybe easier for humans.
• 13:23
  SPEAKER_01: Also,
• 13:24
  SPEAKER_01: I think it also depends on which kind of photo interpreter you are,
• 13:29
  SPEAKER_01: because at Hygiene,
• 13:30
  SPEAKER_01: very high experts are able to distinguish different kinds of trees,
• 13:34
  SPEAKER_01: which is still a very hard task for machines,
• 13:37
  SPEAKER_01: for instance.
• 13:39
  SPEAKER_01: But for land cover,
• 13:40
  SPEAKER_01: I think it will be more about,
• 13:41
  SPEAKER_01: for machines,
• 13:42
  SPEAKER_01: especially about the difference between bare soil,
• 13:45
  SPEAKER_01: agricultural land,
• 13:47
  SPEAKER_01: things like that.
• 13:48
  SPEAKER_00: Right, okay.
• 13:49
  SPEAKER_00: Because presumably a human has additional context.
• 13:52
  SPEAKER_00: Yes, exactly.
• 13:53
  SPEAKER_00: They have a map.
• 13:54
  SPEAKER_00: Okay, that makes a lot of sense.
• 13:56
  SPEAKER_00: You mentioned about adding more modalities.
• 13:58
  SPEAKER_00: I think you mentioned Sentinel-1.
• 14:01
  SPEAKER_00: Presumably that's widely available.
• 14:03
  SPEAKER_00: But is there kind of a wish list of modalities that you would like to include in the future?
• 14:09
  SPEAKER_01: We also think about adding what we can call kind of spot images,
• 14:15
  SPEAKER_01: also satellite images,
• 14:16
  SPEAKER_01: but at one meter per pixel.
• 14:19
  SPEAKER_01: So it's like an intermediate between a sentinel at 10 meters and aerial images at
• 14:24
  SPEAKER_01: 20 centimeters resolution.
• 14:27
  SPEAKER_01: We also think about adding LiDAR information,
• 14:33
  SPEAKER_01: because for now,
• 14:35
  SPEAKER_01: so iGen,
• 14:35
  SPEAKER_01: we're also working on LiDAR acquisition for different land descriptions,
• 14:42
  SPEAKER_01: for instance,
• 14:42
  SPEAKER_01: like forestry mainly,
• 14:44
  SPEAKER_01: but not only,
• 14:44
  SPEAKER_01: also flow estimation.
• 14:48
  SPEAKER_01: And so we think that adding this kind of information,
• 14:50
  SPEAKER_01: like LiDAR,
• 14:50
  SPEAKER_01: which is a high-resolution LiDAR,
• 14:52
  SPEAKER_01: like 10 points per meter square,
• 14:55
  SPEAKER_01: could be interesting to add on the same region and to see if we are able to learn
• 14:59
  SPEAKER_01: models that can deal with different modality or try to leverage the different
• 15:04
  SPEAKER_01: information in the different modality.
• 15:06
  SPEAKER_00: That's a good point to segue back to the model side a bit.
• 15:10
  SPEAKER_00: as you add modalities you add number of channels does that indicate that different
• 15:14
  SPEAKER_00: kinds of models would be successful as you kind of increase the dimension
• 15:18
  SPEAKER_01: dimensionality in that way uh what we observe is the fact that for instance on a
• 15:24
  SPEAKER_01: let's say on the simplest case uh so in our case for the hourly information we have
• 15:29
  SPEAKER_01: of what we call five channel we have rgb infrared and also an elevation model that
• 15:36
  SPEAKER_01: we compute from photogrammetry based on the our images
• 15:40
  SPEAKER_01: And we have shown that adding the infrared increases the performance and also
• 15:46
  SPEAKER_01: adding elevation can increase performance.
• 15:49
  SPEAKER_01: So our clue is the fact that adding more modality,
• 15:52
  SPEAKER_01: if you do it on the right way,
• 15:54
  SPEAKER_01: can bring better performance.
• 15:56
  SPEAKER_01: Okay.
• 15:57
  SPEAKER_00: So that's even without changing the architecture of the model.
• 16:00
  SPEAKER_00: That is just showing that these features have useful information.
• 16:04
  SPEAKER_01: But I think for time series on the Flare dataset baseline we released,
• 16:10
  SPEAKER_01: we didn't add the time series as an extra channel,
• 16:15
  SPEAKER_01: but more like context information that will be processed by an auxiliary branch and
• 16:21
  SPEAKER_01: then injected in the feature space of the main model.
• 16:26
  SPEAKER_01: We did more like that.
• 16:29
  SPEAKER_00: You mentioned briefly to me offline that you saw the benefits of using transformer models.
• 16:35
  SPEAKER_01: Yeah,
• 16:36
  SPEAKER_01: on the first challenge,
• 16:37
  SPEAKER_01: we saw that the best performance was obtained mainly based on a vision transformer,
• 16:46
  SPEAKER_01: which like we can see in the other field of computer vision,
• 16:49
  SPEAKER_01: like in medical case,
• 16:50
  SPEAKER_01: natural images.
• 16:51
  SPEAKER_01: So it was not a surprise,
• 16:53
  SPEAKER_01: but it's interesting to see that some people were able to train this model on our
• 16:56
  SPEAKER_01: data set and to get very good results.
• 16:58
  SPEAKER_01: So now we're trying to incorporate that in our own pipeline.
• 17:03
  SPEAKER_01: and also to release a pre-trained model based on Vision Transformer.
• 17:08
  SPEAKER_01: One of the main problems is the fact that it's sometimes more costly.
• 17:12
  SPEAKER_01: And as I mentioned before,
• 17:13
  SPEAKER_01: as we are trying to produce large-scale data,
• 17:17
  SPEAKER_01: it could be a very clear problem for us.
• 17:21
  SPEAKER_01: And on the other hand,
• 17:22
  SPEAKER_01: some of the problems are the fact that Vision Transformer are based on patches from
• 17:26
  SPEAKER_01: the images.
• 17:27
  SPEAKER_01: Sometimes you can have kind of border effects on the patches.
• 17:31
  SPEAKER_01: That could be a problem for us also, and we try to resolve.
• 17:36
  SPEAKER_01: So maybe more model-like hierarchical vision transformer could be a solution.
• 17:41
  SPEAKER_00: Fascinating stuff.
• 17:43
  SPEAKER_00: Well, thanks so much for this introduction to the dataset series and the accompanying modeling work.
• 17:49
  SPEAKER_00: You mentioned that the third dataset will be out.
• 17:51
  SPEAKER_00: Do you have a date in mind for that?
• 17:55
  SPEAKER_01: Not yet, but I think it will be in 2025.
• 17:59
  SPEAKER_00: Okay, so early next year, hopefully.
• 18:02
  SPEAKER_00: Good stuff.
• 18:03
  SPEAKER_00: If people want to follow along,
• 18:05
  SPEAKER_00: updates,
• 18:05
  SPEAKER_00: both yourself and the activities at IGN,
• 18:08
  SPEAKER_00: which is the best platform?
• 18:11
  SPEAKER_01: Uh, it depends for the, um, more like the innovation part and all the AI model.
• 18:17
  SPEAKER_01: We are hosting a few data set and models.
• 18:21
  SPEAKER_01: Um,
• 18:21
  SPEAKER_01: again,
• 18:21
  SPEAKER_01: the face,
• 18:22
  SPEAKER_01: we have a page from IGN that then can share,
• 18:25
  SPEAKER_01: uh,
• 18:26
  SPEAKER_01: and for all the pollution parts,
• 18:29
  SPEAKER_01: uh,
• 18:29
  SPEAKER_01: it's the IGN website,
• 18:30
  SPEAKER_01: uh,
• 18:31
  SPEAKER_01: where we propose all the information I mentioned by about biodiversity,
• 18:35
  SPEAKER_01: about,
• 18:36
  SPEAKER_01: uh,
• 18:36
  SPEAKER_01: forestry,
• 18:37
  SPEAKER_01: uh,
• 18:37
  SPEAKER_01: agricultural information.
• 18:39
  SPEAKER_01: that are produced by all the institutes.
• 18:42
  SPEAKER_00: Okay, fantastic.
• 18:43
  SPEAKER_00: I'll put the links in the show notes.
• 18:45
  SPEAKER_00: Once again, Nicolas, thanks for joining me and I wish you all the best for the future.
• 18:49
  SPEAKER_00: Thank you.
• 18:50
  SPEAKER_00: You too.
• 18:51
  SPEAKER_00: Bye.
• 18:52
  SPEAKER_00: Bye.