How is Fake Aerial Imagery Created?
In order to create fake imagery, first, you must obtain a great deal of real imagery. These images will be used to train your neural network to recognize the unique landscape characteristics that make a city a city. The machine learning models learn to identify and recreate the patterns we associate with small-scale views of cities (i.e Landsat or Sentinel imagery) mimicking the transitions from city center, to suburbs, to farmlands. They can even be trained to create fake DEMs. ThisCityDoesNotExist.com’s algorithm actually involves creating two competing machine learning algorithms, the generator, and the discriminator. The generator algorithm generates images, attempting to fool the discriminator model which dubs each image as a ‘real’ or ‘fake’ city. Based on the responses of the discriminator model, the generator will adapt its next image, and try again. In programming, this technique is called a Generative adversarial network (GAN). GANs are especially powerful algorithms, and can even be pit against other algorithms to expose their weaknesses. The positive feedback loop the generation and discrimination process creates can hypothetically go on forever, but will ultimately be limited by computer resources, or be influenced by the finite number of images originally used to train the model. These algorithms use unsupervised classification, which is when the computer groups like pixels into a prescribed number of classes. Unsupervised classification distinctly does not require drawing bounding boxes around features, and therefore may be faster to prepare than supervised methods. It is possible to add some influence to the process with supervised classification. This involves inputting a dataset of city center points, telling the model “here, these are cities”, potentially improving results.Why Create Fake Imagery?
In a data rich world with so much access to all varieties of raster data, why would we want to add fake imagery into the mix? Well, the truth is we aren’t quite sure yet. When plenty of real data exists, there is little desire to inject energy into simulating data. While there are no especially compelling uses for large amounts of fake imagery, doctored imagery is a different story. There may be military applications for faking parts of images, but not necessarily whole landscapes. This could be identifying and disguising sensitive military operations in satellite imagery, or maybe even simulating the result of damage to landscape elements. Randomly generated AI landscapes don’t have much use on their own, but once you add more channels to the data (Elevation models, population density, etc), and start adding constraints, things can get really interesting. The trick here is the channels must be coregistered to be meaningful. Coregistering means that the images have all been lined up with each other spatially. By combining information from many channels, you can create a living modeling environment. It gets even more exciting if you can pull off adding time. This allows predictive simulation of urban planning scenarios, deforestation patterns, and even of events resulting from climate change. By simulating different population counts and densities, we can begin to identify common patterns in how people grow into their cities. Fun Fact: If you drop oat flakes into slime mold at the positions of train stations in Tokyo, the resulting fungal network grows to look just like the actual Tokyo Rail System. The reason landscape modeling does not qualify as a compelling application of fake imagery is that it is incredibly costly in terms of computing resources. Without adequate financial backing, it does not make sense to create a program like this. For now, coders are just having fun without funding.What Are the Risks of Fake Satellite Imagery?
The most obvious risk associated with creation of imagery from machine learning models is the risk of people using this technology maliciously. Deep faking images may allow you to represent extreme scenarios on a landscape. One popular one is predicting the same landscape in all four seasons, but you could also simulate flooding, or wildfires. Luckily there is not much genuine risk of misinformation (amongst free media at least) about events as these things are so easily verifiable. If someone actually takes the time to simulate media coverage of a catastrophic event somewhere using deep-faked images, fact-checkers can verify from half a dozen public imagery sources if the event actually occurred (or even simply just call someone there). So if you come across a fake image in the wild, how would you be able to tell it was fake? Well, one characteristic is that fake imagery often looks downscaled. Depending on the base logic in the algorithm used, sometimes you can pick out a fake image by the lack of truly straight road segments. This also applies to the lack of natural angles on building corners, or the presence of especially blurry building footprints. Characteristics like haze or odd colors in areas are not reliable tells of fake imagery as these elements are normal in most real imagery as well due to atmospheric effects. Overall, fake imagery has the potential to be very powerful, but in the current technical landscape, there is not much need for it. Due to the high skill, time, hardware, and financial investment necessary, it is unlikely we will see too much development growth in this area. The one thing we have learned from artificial intelligence, however, is to expect surprises.In Conversation
From Geographer to AI-and-EO Founder
Daniel: Ron, welcome to the podcast. Can you introduce yourself and tell us how you got involved in geospatial?
Ron: I’m Ron, founder and CEO of Osir.io, a small company specialised in AI and Earth observation. My background is in geography. I found an indirect way into remote sensing during my bachelor’s — I got more and more interested in data-heavy topics, originally geo marketing and data mining. In Germany the social and physical sciences are pretty split, so the methods led me from one side to the other. I did my PhD on deforestation in the Amazon, lots of land cover mapping and conventional topics, and over the years got more interested in experimental applications.
Daniel: What was geo marketing?
Ron: That was more than 10 years ago — a small seminar study about segmenting customers and clients. Nothing to do with remote sensing, but the methods — clustering, k-means, old-school data mining — were interesting. Following that into remote sensing was a straight line: classifiers, neural networks, and now generative methods.
ThisCityDoesNotExist.com
Daniel: I found you through thiscitydoesnotexist.com. Tell me about it.
Ron: I was surprised to see people still referencing it — my friend Thomas Langner and I made it nearly two years ago. The most famous of these sites is thispersondoesnotexist.com, and there’s a whole list at thisxdoesnotexist.com of generated images. We reused the StyleGAN method to create Earth observation data and made it in the same style. We’ve also generated other things — not just city imagery but DEMs. The data on the website is Sentinel-2, but we used other data sources for other generators.
How a GAN Actually Works
Daniel: What’s a GAN?
Ron: A generative adversarial network is two neural networks. The discriminator’s job is to decide what’s a real image and what’s a fake. The generator starts from random noise — usually sampled from a Gaussian distribution — and produces gibberish at first, but it’s trained to reduce the accuracy of the discriminator. In their interaction, the generator gets better at fooling the discriminator while the discriminator gets better at telling real from fake. If you balance the training well, you can produce images that look fairly OK — as a human you can still tell, but they’re aesthetically close to real images.
Daniel: Does this just keep going forever?
Ron: It converges depending on the complexity of the network, and it’s limited by the amount of input data. It’s computationally heavy — we used thousands of remote sensing images. Two things about EO make this both interesting and challenging: satellite images are huge — often 10,000 by 40,000 pixels, which kills any memory budget, so we worked with chunks (up to 1024×1024) and typically three or four channels. On the inviting side, it can be unsupervised — for ThisCityDoesNotExist we extracted Sentinel-2 patches by pulling city centre coordinates from OpenStreetMap and threw them in for training. No annotated data, no expert knowledge needed.
Multi-Channel and Vector Possibilities
Daniel: Could I add elevation, population layers — as many channels as I want?
Ron: Yes, limited by memory. Cross-correlations between layers might have positive or negative effects depending on what you’re trying to achieve, but as long as the layers are well co-registered there’s no fundamental limit.
Daniel: What about vector data?
Ron: Yes — I’ve been thinking about this. A lot of urban data is in vector form, often graph-structured — road networks, utility networks. Generative methods are used in medicine to create new molecules, and the data source is the same kind of graph. You could imagine adapting them to create new road networks conditioned on terrain, EO data, whatever, and potentially get more efficient city planning concepts. It’s an idea I’ve had once or twice — never experimented with it — but the methods exist and aren’t limited to images.
Conditioning on Population Density
Daniel: Tell me about the population density example you showed me.
Ron: We did a follow-up project — published in Europa’s Climate Change — using a concept called conditioning. The generative methods learn what structures look like for a given input layer. We took the JRC global population density map (about 1 km resolution) and trained the network on how Sentinel-2 imagery patterns relate to population density. Thomas built a tool where you draw a population density and click a button — the network generates a reasonable-looking image. What’s most impressive isn’t just where you put the brush — it fills the gaps, it figures out how the city should sprawl, agriculture appears around urban centres, and if you draw two separate population peaks they grow together naturally as you increase them. That reasoning is the most interesting part.
Spotting a Fake Satellite Image
Daniel: Are there artifacts that let us spot a generated image?
Ron: The most obvious giveaway in what we produced is that roads aren’t very straight. Thomas would be more qualified to say why — StyleGAN was originally developed for faces, which lack straight structures. So no matter how good the training, you get slightly blurry, curvy lines where humans build very cornered streets. Some “false friends” people point at — strange shades, atmospheric haze — are real artifacts of EO data and not signs of a fake. The real pointer is curvy roads.
Super-Resolution and Cloud Filling
Daniel: Where does super-resolution fit in?
Ron: The methods are similar. You’re generating information at a sub-pixel level. You feed the network a lower resolution image and train against a higher resolution reference. The network learns to produce higher resolution pixels based on the correlations between the two resolutions. If you have two sensors — one high resolution, one lower — once the model is trained you can ignore the high-res data and upgrade the low-res data directly. There are lots of discussions about accuracy and false correlations, but that’s the broad approach.
Daniel: Could this fill missing data in imagery — cloud masking?
Ron: Yes — research shows these methods are used for cloud filling and gap-filling in computer vision more broadly. It does need validation: if you’re going to do a crop classification, sometimes it’s easier to fill the analytics product than the image. Filling reasonable image data is often overkill — you get harmonious data, but the real value would come from extending the methods to include time series information, which doesn’t exist out of the box.
Deep Fakes, Adversarial Networks, and EO’s Future
Daniel: Could someone create fake wildfire imagery by style-transferring smoke into clean scenes?
Ron: Yes — there’s a concept called style transfer where components of one image get overlaid into another and it all looks scarily realistic. In EO I’ve seen it for season changes — summer to snow. You could put smoke into a scene if you have a good data set. But for me, the lower-effort form of fake news is just taking real images out of context. With Ukraine, my LinkedIn feed is full of satellite images, and “open source intelligence” gets thrown around — but the acquisition itself is selective. Images don’t get constantly acquired by sensors, they’re ordered. Distribution is selective too. There’s no truly transparent archive of high-resolution imagery, just the lower-resolution ESA and NASA data. Just publishing a real JPEG out of context already creates a lot of damage. And training a generative method on very high-resolution data would require a lot of proprietary material — a real blocker for the average bored EO scientist.
Daniel: Could the discriminator side of the algorithm be used to detect fakes elsewhere?
Ron: Yes — there’s a lot of discussion in AI about transferability. Things transfer reasonably but still need adaptation. You could have multiple detectors trained on different sensors and a variety of potential fakes — it’s not out of the box, it’d require a lot of man-hours, but it’d be highly useful, especially for investigative data-driven journalists wanting to ensure imagery integrity.
Ron: The other relevant angle is fooling networks rather than detecting fakes. You’ve seen people avoiding face detection with patterns on their glasses, or Google demonstrating a specifically shaped object that becomes invisible to a car’s lidar. In EO that could become relevant for insurance fraud or — without opening the can of worms — military applications, where you hide things from automated ML-based extraction. As more EO data is processed on-edge in space and only analytics products are sent down, dependence on ML grows. If the ML can be fooled, the fun starts. The proportion of EO data that any human eye ever sees will become smaller and smaller, even as the absolute amount grows — so robustness and trustworthiness of the ML becomes a huge topic.
Daniel: Where can people connect with you?
Ron: Thanks again to Thomas Langner, my collaborator on most of this. I’m not very present on social media — you can find me on LinkedIn by name, or drop a mail to ron@osir.io. The easiest is just to go to thiscitydoesnotexist.com — there’s a poster linked at the bottom with contact info and a Colab notebook to play with.
