Transcript

‍

JD (Johannes Dienst): In episode 2 we want to focus on geospatial analysis. What is that, Johannes?

JH (Johannes Haux): Well again, that's in terms of computer vision, something where we look at images, but in this case we're not looking at images of cats and dogs, but of images that potentially have been taken by satellites from the earth, from above basically.

JD: But of cats and dogs. Either way, are you a cat or a dog?

JH: I'm a cat person all the way. But I love dogs, so just that I wouldn't want to take too careful one at my home. Our cat probably wouldn't approve. Yeah, but of course, you know, you can use satellites to detect dogs or cats, so there might be applications there that I'm not aware of.

JD: Okay, what can we do with satellite data, for example?

JH: Well, you can do a lot.As we talked about last episode, computer vision applications are usually, like most applications, trying to make our lives easier to solve our tasks that involve looking at this data.There are a few prominent examples, I guess.For example, weather forecasts, right?That's where we work with geospatial data.I know where the clouds are.I know where there's high air pressure, low pressure fields, where's precipitation, where's the air moist.This kind of stuff. And I can use this data to build models.Usually you don't do this as a computer vision task, but I've seen approaches where you tryto take the two-dimensional projection of this data to make short-term forecasts.

JD: Like a few hours for the rain that comes down.

JH: Yeah, or let it be like 15 minutes. That can also be of importance, right?

JD: These are actually pretty accurate. I always use this when I go for a walk.It just rains outside.

JH: Oh, you do?

JD: Yeah, it works.

JH: So I don't know if there are already real-world applications where this is being done, butI would guess that's a very sensible thing.Because usually what's also a consideration there is compute.So how much energy do I have to spend to come up with a prediction that is accurate enough that I can show it to people, right?And those huge weather models, they are quite expensive, but they are the best thing we can do.Like a true simulation to make forecasts.But if I just want to interpolate between two time steps, for example, I don't need the entire world being modeled. Maybe don't even need the, like only Germany, for example, but I only need this for a very small location.And there it would make sense to have a model that can sample from some expected data distribution between two points.And that's where neural networks are actually pretty good at.But that's only one application.There are more applications you could think of.One that's also quite in the news these days is forest fire prediction.So I'm looking at forests and I try to figure out like, is there a high risk of a fire breaking out here so that I can do, for example, resource allocation.Where do I position my firefighters, et cetera.

JD: Another approach or another task we can solve with computer vision for geospatial data is if there are natural disasters, for example, and I need to get relief in the area of the catastrophe, like how do I get there?And what's the most efficient way to get the needed goods to those critical areas and soon.Like these kinds of tasks you can solve.

JH: And then there's of course always the military that's also very interested in these kinds of things.And they always or usually try to figure out the questions of intelligence.Like where are the bad guys at?Where do I need to drop my bombs?These kinds of things usually not or sometimes from satellites, I guess, but also oftentimes using drones.And then you have like an image of a desert and a bunch of rocks.But one of those rocks is a bad guy I want to take out.So again, computer vision can be used to help people.But of course, like many tools, it can also be used to do kind of ethical, questionable stuff like killing people.

JD: So this would be object detection tasks like for example, find missing person or find a person.

JH: Yeah, find a missing person. Like actually, very, very nice application.The positive sense for finding people in an image would be if you have, for example, a ocean and it's really, really hard to distinguish waves, for example, like the white crowns of a wave from parts of a plane, also white usually.Those are large usually.And then think of finding a person in an ocean.So there again, having computer assist a tool that can scan large amounts of data is really, really helpful.

JD: Could we also use to track change over time?Like for example, the ice melting, the glaciers.

JH: Yeah, that's being done.So the question there would be what's the questions I want to ask about this change.The documentation I would say is there for many areas of the world.And now where you could use deep learning, especially would be to make predictions about how those changes might continue.Or you could again, try to figure out like what were causes for accelerations and change and correlate with other kinds of data like temperature, CO2 level in the atmosphere and so on.

JD: That sounds quite interesting.Recurrent moment.There's also another application.I don't know if it's falling into geospatial analysis, to be honest.There's this object reconstruction and I think it was a few years ago was an ad fromMicrosoft for their eye platform where someone went into ruins and just scanned the image or scanned the surroundings and then they reconstructed the whole place virtually.Is that geospatial analysis?

JH: I mean, probably not, but it's a really interesting application.So for me, geospatial means somehow concerned with the earth, right?And spatial.So some kind of question about how are things related in a spatial sense or so, like looking at a map and reconstructing buildings, of course, means that I need some kind of building plan.So there is this spatial component to the task, but I would guess it's more like structure from motion thing where I, as I said, like walk around, take photographs from various angles so that I can figure out the spatial dimensions of the building and then I can make assumptions of what this building might have looked like.And then I can do an AR application which renders this to me so that I can look at it, which is, to be honest, that sounds really amazing.I would have loved to do that myself.

JD: Cool.So when I think about insurers, for example, this is very famous.So in the city, you pay more for insurance, at least it's Germany, and also in the specific neighborhoods you pay more.Is that an application for geospatial analysis?

JH: I mean, of course.So property value estimation is being done automatically.I would say these days, I'm not totally familiar with the field, but it's definitely something where you use geospatial data.To be honest, the computer vision perspective on this topic I'm not really familiar with, but what you can do, for example, ask questions about property size or population density.And that usually correlates really strongly with property values and with then, of course, also the prices you have to pay for insurance.

JD: Basically, a risk analysis.

JH: Yeah.And if you then combine this with disaster prediction, where will be the next flash flood?Where will be the next mountain coming down?That of course factors into that probably as well.

JD: So you can say, okay, if that's built there, it's a high risk, it will get destroyed at some time.

JH: Yeah, or if you move into this house, we won't insure you because there's like this huge mountain that's been shifting for the past 10 years.Okay, that's already happened.

JD: Yeah, that's already a case.

JH: I've seen videos now from, I think, Switzerland and Austria, where there are entire villages that have to be evacuated because mountains are moving.

JD: That's really interesting.

JH: At least Austria, from what I know, they have really accurate 3D data of their mountains because they are interested in understanding how these mountains change over time to do these kinds of predictions.And they have really high resolution in terms of spatial, but also in terms of temporal resolution images and 3D scans of their mountain ranges.

JD: Well, that's kind of interesting.

Okay, this marks already the end of episode two.So I think we did not exhaust all the geospatial analysis applications, but I think the oneJohannes knows about.

JH: Yeah, yeah, definitely.

See you in the next episode where we will talk about generative AI and the real world applications.

Podcast Homepage