Latest / Satellite image deep learning / Field boundary detection with Segment Anything

Transcript

Field boundary detection with Segment Anything · Satellite image deep learning

• 0:05
  SPEAKER_01: Hello, I'm Robin Cole and welcome to the Satellite Image Deep Learning Podcast.
• 0:09
  SPEAKER_01: In this episode,
• 0:10
  SPEAKER_01: I caught up with Samuel Bancroft to learn about segmenting field boundaries using
• 0:14
  SPEAKER_01: Segment Anything,
• 0:15
  SPEAKER_01: also known as SAM.
• 0:17
  SPEAKER_01: SAM is a foundational model for vision released by Meta, which is capable of zero-shot segmentation.
• 0:24
  SPEAKER_01: However, there are many open questions about how to make use of SAM with remote sensing imagery.
• 0:29
  SPEAKER_01: In this conversation,
• 0:30
  SPEAKER_01: Samuel describes how he used SAM to perform segmentation of field boundaries with
• 0:35
  SPEAKER_01: Sentinel-2 imagery over the UK.
• 0:37
  SPEAKER_01: His best results were obtained not by fine-tuning SAM,
• 0:40
  SPEAKER_01: but by carefully pre-processing a time series of images into HSV color space and
• 0:46
  SPEAKER_01: using SAM without any modifications.
• 0:48
  SPEAKER_01: This is a surprising result,
• 0:50
  SPEAKER_01: and using this kind of approach significantly reduces the amount of work necessary
• 0:54
  SPEAKER_01: to develop useful remote sensing applications utilizing SAM.
• 0:58
  SPEAKER_01: I hope you enjoy this episode.
• 1:01
  SPEAKER_01: Hi, Sam.
• 1:01
  SPEAKER_01: Welcome to the podcast.
• 1:03
  SPEAKER_00: Hiya.
• 1:03
  SPEAKER_00: Thanks for inviting me.
• 1:05
  SPEAKER_01: It's my pleasure.
• 1:05
  SPEAKER_01: Let's begin with where you work and what you do.
• 1:08
  SPEAKER_00: Sure.
• 1:10
  SPEAKER_00: I'm a PhD student at the University of Leeds.
• 1:12
  SPEAKER_00: I've been doing my PhD since 2020, and it's all been about crop type classification.
• 1:18
  SPEAKER_00: So like the rest of machine learning, you're always on the hunt for more and more labels.
• 1:23
  SPEAKER_00: Crop type classification is no different.
• 1:26
  SPEAKER_00: So I'm trying to develop new machine learning methods that rely less on the labels
• 1:32
  SPEAKER_00: that are very expensive to acquire and just making things that are more
• 1:35
  SPEAKER_00: generalizable so I can...
• 1:38
  SPEAKER_00: do big mapping campaigns and work out where the crops are growing across the whole country.
• 1:42
  SPEAKER_00: Right, so that's the goal.
• 1:44
  SPEAKER_00: Yeah, that's the goal.
• 1:45
  SPEAKER_01: Limited data or annotated data, I know I end.
• 1:48
  SPEAKER_01: A model that would generalize across the whole country, in this case, the United Kingdom.
• 1:52
  SPEAKER_01: Okay, well, where do we begin?
• 1:53
  SPEAKER_01: What solution have you come up with?
• 1:56
  SPEAKER_00: Sure.
• 1:56
  SPEAKER_00: So it's been a long process, but I'm talking to you today about UK fields.
• 2:03
  SPEAKER_00: So that's a data set containing millions of field boundaries across England,
• 2:08
  SPEAKER_00: Scotland,
• 2:09
  SPEAKER_00: Wales and Northern Ireland.
• 2:11
  SPEAKER_00: That has been in collaboration with Jake Wilkins, who I believe has been on this podcast previously.
• 2:17
  SPEAKER_00: So we started a conversation about using Meta or Facebook's new machine learning
• 2:23
  SPEAKER_00: model that came out April last year called Segment Anything.
• 2:26
  SPEAKER_00: And a lot of people have tried to use that to apply to Earth observation.
• 2:30
  SPEAKER_00: And we've applied that to the UK to create field boundaries.
• 2:34
  SPEAKER_01: Right.
• 2:34
  SPEAKER_01: And what are some of the limitations that SAM has?
• 2:40
  SPEAKER_00: Sure.
• 2:40
  SPEAKER_00: So...
• 2:43
  SPEAKER_00: SAM is built to,
• 2:46
  SPEAKER_00: or Segment Anything,
• 2:47
  SPEAKER_00: otherwise known as SAM,
• 2:49
  SPEAKER_00: is built to segment a wide range of many different things from people in
• 2:53
  SPEAKER_00: photographs to cats and dogs and road signs.
• 2:56
  SPEAKER_00: It's not necessarily adapted to aerial imagery,
• 2:59
  SPEAKER_00: although perhaps aerial imagery was part of the original training set.
• 3:04
  SPEAKER_00: It's also built around
• 3:06
  SPEAKER_00: rgb red green blue images and we all know that earth observation images don't
• 3:10
  SPEAKER_00: necessarily need to just be rgb so when we take a big foundation we'll like segment
• 3:17
  SPEAKER_00: anything and we apply it to earth observation we don't necessarily get the results
• 3:20
  SPEAKER_00: we want so we want to do feature engineering or fine tuning and stuff like that to
• 3:24
  SPEAKER_01: bring it more into our domain interesting so obviously you want to apply sam to the
• 3:28
  SPEAKER_01: the boundary detection what's some of the stages that you went through in your
• 3:31
  SPEAKER_01: experimentation and where did you arrive at
• 3:35
  SPEAKER_00: So the story actually began at a workshop last year that I attended and some
• 3:42
  SPEAKER_00: academics from a university in Germany set out with this initial task of how can we
• 3:49
  SPEAKER_00: apply segment anything
• 3:51
  SPEAKER_00: to field detection.
• 3:53
  SPEAKER_00: And I think my original attempt was to look at things like NDVI and edge detection
• 3:59
  SPEAKER_00: and feed that in,
• 4:00
  SPEAKER_00: stuff like that.
• 4:01
  SPEAKER_00: We also tried to fine tune the model, which I can talk a little about that later as well.
• 4:06
  SPEAKER_00: And there was some success,
• 4:09
  SPEAKER_00: particularly lots of success in like easy agricultural landscapes like the US
• 4:14
  SPEAKER_00: Midwest where the fields are really,
• 4:16
  SPEAKER_00: really obvious.
• 4:17
  SPEAKER_00: But when I applied it to the UK, just for my own personal bias,
• 4:21
  SPEAKER_00: It didn't do well across the board,
• 4:23
  SPEAKER_00: not well enough to outcompete the traditional field boundary detection methods that
• 4:28
  SPEAKER_00: are out there.
• 4:28
  SPEAKER_01: So North America, the fields are very large and irregular grids in the UK.
• 4:33
  SPEAKER_01: They're maybe not so regular.
• 4:35
  SPEAKER_01: Is that the case?
• 4:36
  SPEAKER_00: Definitely.
• 4:36
  SPEAKER_00: Yes.
• 4:38
  SPEAKER_00: Very flat, very big fields in the US.
• 4:42
  SPEAKER_00: across huge areas in the uk it can change very very quickly particularly between
• 4:47
  SPEAKER_01: that urban and rural areas right okay very interesting and what data did you use
• 4:52
  SPEAKER_01: for this or did you restrict yourself to any one data source sure so uh the uk
• 4:57
  SPEAKER_00: fields data set as is released um is built upon sentinel 2 imagery uh mainly
• 5:03
  SPEAKER_00: because it's a framework
• 5:05
  SPEAKER_00: hopefully anyone will want to use for Sentinel-2 and it gets good high quality
• 5:09
  SPEAKER_00: field boundaries at the 10 meter scale.
• 5:12
  SPEAKER_00: That doesn't necessarily mean that you just need to use Sentinel-2.
• 5:15
  SPEAKER_00: We're releasing the code so you can just swap out
• 5:19
  SPEAKER_00: for any other data source.
• 5:23
  SPEAKER_00: So for example, we've been playing around with planet imagery as well.
• 5:26
  SPEAKER_00: It works with some quite interesting results at the five meter scale as well.
• 5:30
  SPEAKER_01: Very interesting.
• 5:31
  SPEAKER_01: Do you want to talk us through the process of using maybe Sentinel to begin with?
• 5:35
  SPEAKER_01: And maybe we can start at the beginning, how you use multispectral data with a three channel model.
• 5:42
  SPEAKER_01: Sure.
• 5:43
  SPEAKER_00: Yes.
• 5:43
  SPEAKER_00: Should I bring up my slides?
• 5:45
  SPEAKER_00: Yeah, go for it.
• 5:46
  SPEAKER_00: Definitely.
• 5:47
  SPEAKER_00: Okay.
• 5:55
  SPEAKER_00: All right, so to begin with the.
• 6:00
  SPEAKER_00: solution to how to make UK fields work consistently across the whole country.
• 6:05
  SPEAKER_00: very much lies on making harmonic composites now this isn't a completely fresh idea
• 6:12
  SPEAKER_00: harmonic composites exist and existed a long time.
• 6:16
  SPEAKER_00: for a variety of different reasons.
• 6:18
  SPEAKER_00: I've used Harmonix to do crop type classification in an unsupervised way,
• 6:24
  SPEAKER_00: so using no labels whatsoever,
• 6:25
  SPEAKER_00: trying to identify distinct different crop types automatically.
• 6:32
  SPEAKER_00: And you can do this in Google Earth Engine relatively easily.
• 6:36
  SPEAKER_00: And other people, no doubt, have been using this with their own field delineation algorithms.
• 6:41
  SPEAKER_01: Just to clarify for the readers,
• 6:42
  SPEAKER_01: the listeners,
• 6:43
  SPEAKER_01: sorry,
• 6:43
  SPEAKER_01: the harmonics are fitted to a time series,
• 6:46
  SPEAKER_01: essentially.
• 6:48
  SPEAKER_00: Yes.
• 6:49
  SPEAKER_00: So what you're looking at here is a little bit of a kaleidoscopic image,
• 6:54
  SPEAKER_00: but it's representing a time series of images.
• 6:57
  SPEAKER_00: And how that's done is it's taking a vegetation index,
• 7:01
  SPEAKER_00: in this case NDVI,
• 7:02
  SPEAKER_00: across a 12-month growing period.
• 7:05
  SPEAKER_00: And it's fitting a harmonic model to that time series.
• 7:09
  SPEAKER_00: And from that harmonic model,
• 7:11
  SPEAKER_00: from each pixel,
• 7:12
  SPEAKER_00: you can get the phase,
• 7:14
  SPEAKER_00: the amplitude,
• 7:15
  SPEAKER_00: and the mean NDVI of that particular pixel.
• 7:20
  SPEAKER_00: And we map that to the HSV hue saturation value color space.
• 7:24
  SPEAKER_00: So the phase, because it's cyclical,
• 7:29
  SPEAKER_00: can be the hue, saturation can be the amplitude, and value can be the mean NDVI.
• 7:35
  SPEAKER_00: And once you pop that and encode that as this kind of static image here,
• 7:40
  SPEAKER_00: you get loads of different colors,
• 7:41
  SPEAKER_00: right?
• 7:42
  SPEAKER_00: And the hope is that something like this is much more easily interpretable by
• 7:48
  SPEAKER_00: segment anything than median RGB composites just in a normal way.
• 7:55
  SPEAKER_01: At a high level, we can see the boundaries are quite clear there.
• 7:59
  SPEAKER_01: I suppose in any one time, just in RGB, the boundaries wouldn't be quite so clear.
• 8:07
  SPEAKER_00: Yes, definitely.
• 8:09
  SPEAKER_00: There's a few good examples about just using an RGB composite and comparing it to
• 8:14
  SPEAKER_00: this and seeing the accuracy jump up a little bit.
• 8:16
  SPEAKER_00: So what we did, actually, is we tested Segment Anything just with normal color imagery.
• 8:23
  SPEAKER_00: compared to the harmonic approach.
• 8:24
  SPEAKER_00: And it bumps up what's known as the dice score, which is a nice accuracy metric.
• 8:30
  SPEAKER_00: It just bumps it up by 0.2 on average.
• 8:34
  SPEAKER_00: So a significant improvement, basically.
• 8:36
  SPEAKER_00: So it really does fix a lot of the problems with a few of the odd fields that don't
• 8:40
  SPEAKER_00: get segmented properly.
• 8:42
  SPEAKER_01: Okay.
• 8:42
  SPEAKER_01: So the basic process is take your time series over 12 months,
• 8:46
  SPEAKER_01: generate one of these HSV images,
• 8:48
  SPEAKER_01: and then pass it through,
• 8:50
  SPEAKER_01: segment anything without any kind of fine tuning or modification for this situation.
• 8:57
  SPEAKER_01: Yeah.
• 8:58
  SPEAKER_00: Fantastic.
• 8:59
  SPEAKER_00: It's a good example of don't reinvent the wheel,
• 9:03
  SPEAKER_00: as perhaps I tried to do when I initially set out to do this and just rely a little
• 9:08
  SPEAKER_00: bit on feature engineering.
• 9:09
  SPEAKER_01: Right.
• 9:10
  SPEAKER_01: Excellent.
• 9:11
  SPEAKER_01: And how do the results, metrically speaking, look across the UK?
• 9:17
  SPEAKER_00: Sure.
• 9:17
  SPEAKER_00: So if I bring up the right slides.
• 9:23
  SPEAKER_00: These are results scoring against the next best available ground truth is validation data.
• 9:31
  SPEAKER_00: So we've taken our UK fields data set.
• 9:35
  SPEAKER_00: We've split it up into the grid cells and worked out the average
• 9:40
  SPEAKER_00: dice score,
• 9:41
  SPEAKER_00: intersection over union,
• 9:42
  SPEAKER_00: under-segmentation and over-segmentation rates compared to the Ordnance Survey.
• 9:46
  SPEAKER_00: So that's the national mapping agency,
• 9:48
  SPEAKER_00: government agency for the UK,
• 9:51
  SPEAKER_00: compared to the fields in the MasterMap dataset that they've got available.
• 9:57
  SPEAKER_00: So for those thinking about harmonic composites already, you'll notice that the fields that
• 10:04
  SPEAKER_00: strongly agricultural that are growing crops that have very distinct growing curves
• 10:08
  SPEAKER_00: will do very very nicely in a harmonic composite and you can see that reflected
• 10:11
  SPEAKER_00: here you can see the east coast uh big growing regions of the uk get very nice
• 10:16
  SPEAKER_00: scores you know 0.7 0.8 dice score which is really really good um when you get to
• 10:24
  SPEAKER_00: uh
• 10:25
  SPEAKER_00: non-agricultural areas like Scottish Highlands or where there's not many crops growing,
• 10:30
  SPEAKER_00: for example Cumbria,
• 10:32
  SPEAKER_00: it doesn't do so well.
• 10:34
  SPEAKER_00: But the hope is UK Fields is releasing so many agricultural fields,
• 10:41
  SPEAKER_00: you're not really very interested in fields
• 10:44
  SPEAKER_00: in non-agricultural areas, right?
• 10:46
  SPEAKER_00: So when we release the data set, there won't be fields over the Scottish Highlands.
• 10:51
  SPEAKER_00: And I don't think anyone's necessarily going to be too worried about the lack of them.
• 10:56
  SPEAKER_01: Yeah, makes a lot of sense.
• 10:59
  SPEAKER_01: Fantastic.
• 10:59
  SPEAKER_01: Do you want to dive into a few more of the details around maybe the pre-processing?
• 11:04
  SPEAKER_01: I think you had a couple of extra slides there.
• 11:07
  SPEAKER_00: Definitely, yes.
• 11:07
  SPEAKER_00: Yeah.
• 11:09
  SPEAKER_00: I should say...
• 11:11
  SPEAKER_00: Releasing this kind of data set is probably borne out a lot of frustration in my
• 11:16
  SPEAKER_00: PhD when I tried to get my hands on field data very easily.
• 11:19
  SPEAKER_00: There's a few offerings in the UK that you can make use of,
• 11:24
  SPEAKER_00: but sometimes their paywalls or the methods aren't very easy to work out compared
• 11:29
  SPEAKER_00: to other countries in mainland Europe,
• 11:33
  SPEAKER_00: which is much,
• 11:33
  SPEAKER_00: much easier to get hold of more timely data.
• 11:37
  SPEAKER_00: And I told myself,
• 11:38
  SPEAKER_00: if I ever made the field boundaries myself for my own research,
• 11:43
  SPEAKER_00: I'd make it available so people wouldn't have to go through the same old slog to
• 11:47
  SPEAKER_00: get the same data.
• 11:48
  SPEAKER_00: So the ultimate aim of this is,
• 11:50
  SPEAKER_00: yes,
• 11:51
  SPEAKER_00: so other researchers can look at fields over the UK,
• 11:54
  SPEAKER_00: but by releasing the code as well,
• 11:56
  SPEAKER_00: We're hoping that other people create more field boundaries over loads of other
• 12:00
  SPEAKER_00: regions all across the world.
• 12:02
  SPEAKER_00: So it's just a single line import in Google Earth Engine and suddenly you've got
• 12:06
  SPEAKER_00: your fields to work with because this kind of data is just a stepping stone for
• 12:11
  SPEAKER_00: next further bits of analysis that you need to do.
• 12:15
  SPEAKER_00: So that's an important thing that hopefully releasing this will make a big difference.
• 12:21
  SPEAKER_00: There's also a few quality,
• 12:23
  SPEAKER_00: useful features in the code release that I hope people can make some good use of.
• 12:31
  SPEAKER_00: We could have released the code and just told you how to segment different tiles,
• 12:35
  SPEAKER_00: because if you're mapping the whole UK,
• 12:38
  SPEAKER_00: you've got to chip it up into lots of different adjacent areas and do your results.
• 12:45
  SPEAKER_00: But when you stitch it all together,
• 12:47
  SPEAKER_00: if it detects one field in one chip and then the next chip over the next adjacent chip,
• 12:52
  SPEAKER_00: it doesn't necessarily get the same field.
• 12:55
  SPEAKER_00: Merging it together becomes a complete headache.
• 12:57
  SPEAKER_00: So what we've got here is what we first do in Google Earth Engine is we export
• 13:03
  SPEAKER_00: overlapping chips,
• 13:05
  SPEAKER_00: probably about 10%.
• 13:07
  SPEAKER_00: either side, we do that in an overlapping fashion such that we can merge them together later.
• 13:13
  SPEAKER_00: And this is just a quick visualization of how we do merge them together.
• 13:16
  SPEAKER_00: So we do like a radial weighting.
• 13:19
  SPEAKER_00: So we take a kernel,
• 13:21
  SPEAKER_00: a radial weighting kernel on the left-hand side that you can see there,
• 13:24
  SPEAKER_00: and we apply that to the inference result we get from segment anything on the
• 13:28
  SPEAKER_00: harmonic composite,
• 13:29
  SPEAKER_00: which you can see in the middle tile.
• 13:31
  SPEAKER_00: Merge them together, and you can say that the results in the very center of that tile
• 13:36
  SPEAKER_00: we're most confident about, and those on the periphery, not so much.
• 13:41
  SPEAKER_00: And then when we stitch together all the overlapping ones,
• 13:44
  SPEAKER_00: so if we go left to right here,
• 13:45
  SPEAKER_00: we take the first tile,
• 13:46
  SPEAKER_00: weight it,
• 13:47
  SPEAKER_00: apply it to the next weighted tile,
• 13:49
  SPEAKER_00: and then the next one,
• 13:50
  SPEAKER_00: and then the next one.
• 13:51
  SPEAKER_00: And it just builds up the weighted average, right?
• 13:53
  SPEAKER_00: Such that if you...
• 13:55
  SPEAKER_00: take all of the weightings, you add them all together, then you binarize it at the end.
• 14:00
  SPEAKER_00: It just means that there's no ugly discontinuities between the different tiles and
• 14:06
  SPEAKER_00: information isn't lost.
• 14:08
  SPEAKER_00: And sometimes this code, when you're doing it at scale, can be a little bit tricky.
• 14:13
  SPEAKER_00: So we're releasing it so you can do this in Google Cloud and the Google Cloud platform.
• 14:18
  SPEAKER_00: So you don't even need to bring any data onto your own laptop to do this kind of stuff.
• 14:23
  SPEAKER_00: And also, if you do want to use stuff on your own laptop, we have that code available as well.
• 14:27
  SPEAKER_01: Right, okay.
• 14:28
  SPEAKER_01: I'm not too familiar with how that would work on Google.
• 14:31
  SPEAKER_01: You just give it the function and when API access.
• 14:35
  SPEAKER_00: Yeah, and there's also some very useful GDAL commands, but...
• 14:41
  SPEAKER_00: it it took me a long time to to get it all in the right place and working the way
• 14:46
  SPEAKER_01: incantation yeah yes awesome that's that's an amazing result and some really nice
• 14:53
  SPEAKER_01: technology built around it to make it very practical for use um i'm also interested
• 14:58
  SPEAKER_01: a bit in the paths that you went down that didn't work out you mentioned uh fine
• 15:02
  SPEAKER_01: tuning of sam this is something that so people have talked about doing but nobody's
• 15:07
  SPEAKER_01: really sure how to do do you mind talking a bit about that
• 15:09
  SPEAKER_00: Yes, no, definitely.
• 15:11
  SPEAKER_00: It's very important to talk about all the failures because if I'd gone through all
• 15:17
  SPEAKER_00: of this and not released it,
• 15:19
  SPEAKER_00: it's been hidden away.
• 15:22
  SPEAKER_00: One of the first big things I tried was fine tuning the Segment Anything to work better on RGB images.
• 15:31
  SPEAKER_00: And I was inspired by, I think it was called MedSam.
• 15:34
  SPEAKER_00: So it was Segment Anything fine tuned against medical images.
• 15:39
  SPEAKER_00: So I took some code off GitHub and tried to play around with it and swapped out the
• 15:44
  SPEAKER_00: medical image datasets for a popular field boundary benchmark dataset called AI for Boundaries,
• 15:52
  SPEAKER_00: which contains loads of Sentinel-2 imagery and corresponding field boundaries
• 15:57
  SPEAKER_00: across a huge part of Europe,
• 16:00
  SPEAKER_00: I believe.
• 16:00
  SPEAKER_00: And I fine-tuned that and
• 16:06
  SPEAKER_00: At first,
• 16:07
  SPEAKER_00: I thought I was getting better results,
• 16:08
  SPEAKER_00: but this is also,
• 16:10
  SPEAKER_00: it was very much just a pet project in the background.
• 16:12
  SPEAKER_00: I was training for many hours at a time, using a lot of GPU time, just playing around with this.
• 16:18
  SPEAKER_00: But the end result was no matter what type of model I use,
• 16:22
  SPEAKER_00: there's three different versions of Segment Anything with different numbers of parameters.
• 16:26
  SPEAKER_00: No matter what I tried, it didn't work.
• 16:30
  SPEAKER_00: it didn't properly bump up the accuracy in any meaningful way.
• 16:33
  SPEAKER_00: In many cases, it made it worse.
• 16:35
  SPEAKER_00: Now, I'm not saying that it's impossible to fine-tune it.
• 16:39
  SPEAKER_00: I'm just saying that my attempt at doing it in very much a background fashion just
• 16:45
  SPEAKER_00: didn't work out in the end.
• 16:46
  SPEAKER_00: It was a lot of work for very little gain.
• 16:49
  SPEAKER_00: And this harmonic composite approach was a lot less work for a lot more gain and a
• 16:56
  SPEAKER_00: lot more consistent results across the board.
• 16:59
  SPEAKER_01: Yeah,
• 16:59
  SPEAKER_01: I think for me,
• 16:59
  SPEAKER_01: that's a standout sort of takeaway that,
• 17:02
  SPEAKER_01: you know,
• 17:03
  SPEAKER_01: you can fine tune models or maybe you can take existing models and experiment with
• 17:06
  SPEAKER_01: the feature engineering,
• 17:08
  SPEAKER_01: as you put it,
• 17:09
  SPEAKER_01: and potentially that's a more viable route.
• 17:13
  SPEAKER_00: Yeah, definitely.
• 17:14
  SPEAKER_00: Sometimes there comes a point where if you've got something that works, you've just got to release it.
• 17:20
  SPEAKER_00: You've got to make sure that other people can build on that as well.
• 17:23
  SPEAKER_01: Really exciting stuff.
• 17:24
  SPEAKER_01: I think on your last slide, you've got the GitHub link.
• 17:26
  SPEAKER_01: Maybe we just put that one up and then we can just say,
• 17:30
  SPEAKER_01: okay,
• 17:31
  SPEAKER_01: so everything's being published at the moment.
• 17:33
  SPEAKER_01: You're finishing your PhD now.
• 17:36
  SPEAKER_01: What's next for you and what are your ambitions after having done this?
• 17:42
  SPEAKER_00: So I came into my PhD from industry and very much at the end of my PhD,
• 17:47
  SPEAKER_00: I'll be looking to return to industry.
• 17:49
  SPEAKER_00: I'm all about agriculture, earth observation, machine learning.
• 17:53
  SPEAKER_00: So I'll be looking for a job in those kind of fields very soon.
• 17:58
  SPEAKER_00: My PhD finishes at the end of October.
• 18:02
  SPEAKER_00: That's October 24.
• 18:02
  SPEAKER_00: So only a few months away.
• 18:05
  SPEAKER_01: Exciting.
• 18:06
  SPEAKER_01: And if people want to follow along your updates, which is the best platform for that?
• 18:11
  SPEAKER_00: uh best uh way to be best informed when this is all released is to look at my
• 18:15
  SPEAKER_01: twitter so that's at spirit well okay fantastic once again thank you uh fascinating
• 18:21
  SPEAKER_01: conversation and all the best for the write-up and hopefully we'll hear from you
• 18:24
  SPEAKER_00: again in the future lovely thank you very much for your time rogan bye cheers