The most dangerous satellite deepfake may not be a completely fabricated landscape. It may be a mostly authentic image in which one small but consequential feature has been added, removed or moved: an aircraft on a runway, a new building, a damaged bridge or a vehicle near a border.
That is the problem addressed by a new prototype benchmark from researchers at Oak Ridge National Laboratory. Their preprint, published on 5 August 2026, introduces a dataset designed not only to classify satellite images as authentic or manipulated, but to identify the exact pixels that were changed.
The timing is striking—and Geoawesome has already documented how quickly the issue moved. On 30 July, we introduced the feature in “Nano Banana Lands in Google Earth and It’s Totally #Geoawesome”, focusing on its potential for planning concepts, historical reconstructions and place-based creative work. Within a day, Google rolled it back after users produced plausible-looking scenes of attacks, disasters and sensitive infrastructure, as we explained in “Google Shuts Down Nano Banana in Google Earth After Misuse Concerns”. The short-lived experiment showed how quickly synthetic content can borrow the visual authority of an established geospatial platform. The Oak Ridge work addresses a narrower technical problem, but one likely to become essential: how should forensic systems evaluate a realistic alteration embedded within genuine remote-sensing data?
Why localization matters
Most deepfake detection is framed as a binary decision. A model receives an image and returns a probability that it is real or synthetic. That may be useful for fully generated pictures, but it is inadequate when only part of an image has been changed.
For a geospatial analyst, the location of the manipulation is often more important than the image-level label. A suspicious object occupying a few dozen pixels could alter an assessment even though more than 99% of the image remains authentic. A localization system should therefore produce a mask showing the manipulated region, giving a human analyst something interpretable to inspect.
The Oak Ridge dataset, called fmow-fake-small, contains 60 images: 30 authentic and 30 manipulated. Every manipulated example has a pixel-level ground-truth mask. The dataset also preserves georeferencing and acquisition metadata, allowing researchers to examine whether forensic performance varies with sensor, pixel size or collection conditions.
That combination is unusual. Large remote-sensing deepfake datasets already exist, including RSFAKE-1M, which contains one million real and one million synthetic examples. But the Oak Ridge researchers argue that existing collections often lack masks suitable for localization, use fully generated images rather than localized alterations, or contain visible artifacts that make detection unrealistically easy.
How the fake images were built
The source imagery comes from Functional Map of the World, or fMoW, a dataset of more than one million images collected across roughly 200 countries by WorldView-2, WorldView-3, GeoEye-1 and QuickBird-2. The researchers used RGB, eight-bit pansharpened images and converted them into georeferenced GeoTIFFs using the accompanying metadata.
They then created ten examples for each of three manipulation types.
The first is a simple splice: a rectangular crop from one image is inserted into another. The benchmark includes splices from 16 by 16 to 256 by 256 pixels. Crucially, the crop is resampled using the physical ground dimensions and pixel resolution of the destination image. This reduces the scale inconsistencies that can reveal a naive cut-and-paste operation.
The second type is an object splice. The researchers used Meta’s Segment Anything model to extract objects or land-cover features and then manually curated the masks and placed the objects in plausible locations. Targets include vehicles, pools, buildings, runways, helipads, tennis courts and areas of vegetation or bare ground.
The third type uses diffusion-model inpainting. The team employed RSPaint, a Stable Diffusion model fine-tuned for remote-sensing imagery. A reference object or surface—such as an aircraft, building, dirt road, agricultural field or green space—is inserted into a user-defined part of the base image.
Here, geographic scale becomes a forensic issue. A generative model may produce a visually convincing aircraft that is physically far too large relative to the runway. The researchers address this by matching the real-world dimensions of reference objects and destination masks. They also crop the working image so the masked region occupies 15% to 30% of the model input, a range that produces more reliable inpainting.
These details make the examples harder than generic AI-image tests. Remote-sensing forgeries must respect not only visual texture but ground sampling distance, object dimensions and the spatial relationships among features.
A useful benchmark, not a finished solution
The dataset’s name is appropriately candid: it is small. Thirty manipulated images cannot support the training of a robust detector, and the authors explicitly recommend using the collection for evaluation rather than training.
Its diversity is also limited. Only one manipulation category uses a generative model, and all diffusion examples come from the same RSPaint workflow. Object placement and mask definition require substantial manual work. This improves visual quality but makes the construction process difficult to scale.
The paper also does not present a comprehensive leaderboard showing how current forensic models perform. It describes a dataset-construction method and releases a prototype benchmark. Claims that it “solves” satellite deepfake detection would therefore be premature.
Yet its limitations reveal an important reality: constructing a small number of credible, spatially consistent satellite forgeries can be more valuable for stress-testing than generating millions of obviously artificial images. A detector that performs well on low-quality fakes may simply be learning compression artifacts, warped geometry or model-specific signatures. Such shortcuts often fail when the image generator changes.
The stronger test is whether a system can identify a small, semantically plausible change without relying on an obvious visual flaw.
Detection is only one layer of trust
Even a much larger version of fmow-fake-small would not be sufficient by itself. Detection models face an adversarial and constantly changing problem: new generators appear, editing methods improve, and common forensic traces disappear through cropping, recompression or screenshots.
Geospatial verification will therefore need several complementary layers. Pixel-level forensic analysis can flag suspicious regions. Acquisition metadata can be checked for internal consistency. Analysts can compare the image with earlier or later collections, alternative sensors and independent providers. Cryptographic provenance could help verify the origin and processing history of imagery before it reaches a public platform.
The most resilient workflow will combine these methods rather than assume that a single AI detector can serve as a universal truth machine.
That is particularly important because satellite imagery occupies an unusual position. It is both a technical dataset and a persuasive visual artifact. Governments, journalists, insurers, environmental monitors and courts use it as evidence about events that may be inaccessible from the ground. A small alteration can therefore carry consequences far beyond the number of pixels involved.
The Oak Ridge benchmark is preliminary, but it asks the right question. In the era of synthetic geography, verifying an image is not enough. Analysts also need to know exactly where reality may have been edited.
The dataset is publicly available through Hugging Face.
#GeoAI
Next article
Satellogic’s First Operating Profit Shows Where the Earth Observation Business Is Heading
Satellogic has reached a financial milestone that commercial Earth observation companies have often promised but rarely delivered: positive operating income. In the second quarter of 2026, the company reported $15.9 million in revenue, $0.3 million in operating income and $2.8 million in adjusted EBITDA.
That does not mean Satellogic has suddenly become unambiguously profitable. It recorded a $20 million net loss for the quarter, and its free cash flow remained negative. But the result is still strategically important because it offers an early test of a changing business model in Earth observation. Satellogic is no longer relying only on selling images or subscriptions. It is combining imagery, monitoring services, AI applications and the transfer of operational satellites to sovereign customers.
The question for the wider geospatial industry is whether this mix can become a repeatable business rather than a collection of large, irregular contracts.
What the numbers actually show
According to Satellogic’s second-quarter results, revenue increased 259% from $4.4 million a year earlier to $15.9 million. The increase came from two different parts of the company.
Data and Analytics, which includes imagery and Constellation-as-a-Service, generated $7.1 million, up from $4 million. Space Systems generated $8.8 million, compared with only $0.5 million in the same quarter of 2025. In other words, more than half of quarterly revenue came from spacecraft-related activity rather than the recurring sale of data and analytics.
That distinction matters. A satellite transfer can produce a large revenue event, but it is not necessarily as predictable as a multi-year data subscription. Satellogic says its remaining performance obligations reached $80.7 million at the end of June, of which $45.8 million is expected to be recognized within one year. This provides contracted visibility, although it does not eliminate delivery, timing or customer-concentration risk.
The profitability figures also need careful interpretation. GAAP operating income was positive for the first time, at $0.3 million, while adjusted EBITDA reached $2.8 million. Yet the company’s bottom line remained negative because of a $19.7 million non-cash fair-value charge associated primarily with convertible notes, warrants and earnout liabilities. Net cash used in operating activities was $8.6 million. After capital expenditure and proceeds from an in-orbit satellite sale, reported non-GAAP free cash flow was negative $5.9 million.
Satellogic therefore crossed an operational threshold, not the finish line. It showed that quarterly revenue can cover operating costs under the right contract mix. It has not yet shown that the model produces consistent net income or positive cash flow across multiple quarters.
The peer comparison puts that milestone in perspective. Planet, the largest publicly traded pure-play EO operator, generated $94.2 million in revenue in the quarter ended April 30—almost six times Satellogic’s quarterly sales. Planet was roughly at break-even on adjusted EBITDA, with a $1 million loss, after posting its first full fiscal year of positive adjusted EBITDA in fiscal 2026: $15.5 million on $307.7 million of revenue. Its headline quarterly net loss of $138.9 million looks dramatic, but $106.5 million came from a non-cash warrant revaluation. That is a useful reminder that, for both Planet and Satellogic, operating income, cash flow and adjusted EBITDA currently say more about the underlying business than GAAP net income alone.
BlackSky is a closer comparison in scale and customer profile, although its results show how lumpy this market can be. It recorded $106.6 million of revenue and $0.9 million of adjusted EBITDA in 2025, its second consecutive adjusted-EBITDA-positive year, while still reporting a $70.3 million net loss. Its first quarter of 2026 then swung to a $5.1 million adjusted EBITDA loss as revenue fell to $20.8 million, before management raised its full-year outlook. Satellogic’s positive quarter is encouraging against that backdrop, but one quarter is not yet the same as the multi-year track record that Planet and BlackSky are beginning to establish.
There is also a private-company benchmark worth noting. Radar operator ICEYE says it produced more than €250 million in 2025 revenue and more than €100 million in EBITDA, alongside over €130 million in operating cash flow. Those are unaudited, company-reported figures, and ICEYE’s radar constellation and sovereign-system sales are not directly comparable with Satellogic’s optical business. Still, they show how large—and potentially profitable—the sovereign EO model can become. Against these peers, Satellogic is not yet the sector’s scale or profitability leader. Its more interesting distinction is that over half of its latest revenue came from Space Systems, making it look increasingly like a hybrid satellite manufacturer, operator and intelligence provider.
Sovereignty is becoming a product
The most revealing part of the result is the growing role of sovereign capacity.
In January, Satellogic signed an $18 million agreement to provide Portugal’s CEiiA with two NewSat Mark V satellites for the Atlantic Constellation. The agreement covers in-orbit delivery and a transfer of ownership and operational control. Satellogic reported that the first satellite had been delivered by the second quarter.
The company also signed a separate $12 million agreement to transfer a commissioned satellite from its operational constellation to an unnamed sovereign defense customer. The package includes support for developing independent command and data-processing capabilities. Another international defense customer expanded an initial trial into a one-year monitoring agreement worth more than $18 million.
These deals reflect a shift in what governments are buying. A country may still purchase imagery from a commercial archive, but many now also want assured tasking, national control, local processing and reduced dependence on external providers. The product is no longer just a pixel. It is an operational degree of sovereignty.
This approach sits between two established models. At one end, governments procure data from large commercial constellations without owning the spacecraft. At the other, they fund bespoke national missions that may take years to design and launch. The transfer of an already commissioned small satellite offers a faster route to national capability, albeit with less customization and continued dependence on the supplier for knowledge transfer and support.
For smaller space nations, that trade-off may be attractive. It also gives an EO operator another way to monetize its manufacturing capacity and orbital assets.
From imagery vendor to intelligence platform
Satellogic is simultaneously trying to move up the value chain. Its recent collaborations with SynMax and SpaceKnow are intended to place automated detection and monitoring applications on top of its imagery infrastructure.
Payload’s reporting on the SpaceKnow partnership described the strategy as a move away from one-off imagery sales toward data-as-a-service and application delivery. In this model, Satellogic supplies collection capacity and historical imagery, while analytics partners build products for activities such as asset, maritime or industrial monitoring.
The logic is familiar across the EO market. Imagery is essential, but customers ultimately pay for decisions: what changed, whether it matters and what action should follow. Partnerships allow Satellogic to add domain-specific analytics without building every application internally.
The risk is that “AI-powered intelligence” remains easier to announce than to operationalize. Reliable monitoring requires more than an object detector. It depends on collection consistency, geolocation accuracy, cloud handling, false-positive management, workflow integration and human review. Customers will judge the system on whether it reduces decision time without creating an unmanageable stream of uncertain alerts.
Merlin remains a company promise
Satellogic’s most ambitious claim concerns Merlin, its planned AI-first constellation. The company says Merlin will eventually remap the entire planet every day at one-metre resolution, use ten spectral bands aligned with Sentinel-2 and employ onboard processing and inter-satellite links. The first launch is targeted for October 2026, with full operations expected in the first half of 2027.
These are forward-looking company claims, not demonstrated operational capabilities. Satellogic says the system is fully funded by existing customer contracts, which reduces financing risk, but launch, deployment, calibration and production performance remain to be proven.
The distinction is important because Merlin underpins the company’s proposed transition from known-site monitoring to broad-area detection. Existing constellations can repeatedly inspect selected targets. A system that genuinely combines daily global coverage, one-metre imagery and automated detection would address a different class of problem: finding changes before an analyst has decided where to look.
That would be strategically valuable for defense and intelligence, but also commercially difficult. Capturing the world is not the same as processing it economically, identifying meaningful change or delivering alerts at acceptable latency.
An inflection point—with conditions attached
Satellogic ended the quarter with $112.8 million in cash, a substantially stronger position than it held before its 2026 financing. It also reduced the principal of its secured convertible notes from $30 million at the end of 2025 to $18 million. Those improvements give it room to deploy Merlin and pursue additional sovereign contracts.
However, the company still faces the structural challenges that have defined commercial EO: capital-intensive infrastructure, long government sales cycles, a limited number of large customers and fierce competition from better-capitalized optical and radar providers. Its own annual filing identifies customer concentration, launch dependence, market adoption and the conversion of prospective contracts into revenue as material risks.
The second-quarter result should therefore be read neither as proof that those problems are solved nor as a routine earnings beat. It is evidence that a vertically integrated EO company can reach operating profitability when it sells a combination of data, monitoring and sovereign space systems.
The next test is repeatability. If future quarters are driven by recurring analytics and persistent-monitoring revenue, Satellogic may be demonstrating a durable model for the next generation of Earth observation companies. If performance continues to depend on occasional satellite transfers, the business will remain lumpy even when strategically relevant.
Either way, the direction of travel is clear. The commercial EO market is moving beyond imagery access. Governments and enterprises increasingly want control, continuity and answers—and providers are reorganizing around those demands.
