03 September, 2026

Seeing More: The State of Super-Resolution in Satellite Imagery

Resolution has always been the currency of satellite imagery. The finer the detail, the more a client can measure, classify, and trust. But launching a higher-resolution sensor is a multi-year, multi-million-dollar undertaking, so increasingly, the industry is asking a different question: can we extract more resolution from the imagery we already have?

That's the premise behind super-resolution, and it's a technology Geoimage has been quietly testing, and applying across both free and commercial satellite imagery for the past 18 months. Working closely with our technology partner, SI-A, models have been tuned against the specific characteristics of different sensors, because a technique that performs well on one platform's spectral and spatial signature won't necessarily transfer cleanly to another.

Two Paths to the Same Goal

Super-resolution isn't one technology, it's two quite different approaches that happen to produce a similar-looking result.

Sharpening-based methods

The more established path. These rely on deterministic algorithms, pan-sharpening, deconvolution, edge-enhancement techniques, that combine information already present in the image (such as fusing a high-resolution panchromatic band with lower-resolution multispectral bands) to produce a crisper output. The mathematics are transparent, the results are explainable, and nothing is invented that wasn't already implicit in the source data.

AI-based methods

A different route entirely. Deep learning models, trained on large datasets of paired low- and high-resolution imagery, learn to infer plausible fine detail that isn't explicitly present in the input pixels. The results can be visually striking, often producing outputs that appear to rival imagery from a genuinely higher-resolution sensor.

Not all AI approaches carry the same risk profile, though. Airbus's own Pléiades Neo HD15 product is a useful illustration: rather than a generative model, it uses a supervised convolutional neural network, deliberately avoiding a GAN-based architecture to improve stability and constrain the risk of fabricated detail. The model is trained on simulated image pairs built from extremely high-resolution aerial imagery, and the enhancement factor is capped at 2x (from 30cm to 15cm) rather than left open-ended. Airbus describes it as image restoration rather than artificial enhancement. It's a good demonstration that "AI super-resolution" isn't one technique with one risk level: the underlying model architecture and training methodology matter a great deal to how much you can trust the output.

The Catch

That last point is exactly where the conversation gets more interesting, and more cautious.

Because AI super-resolution models are inferring detail rather than measuring it, there's an inherent risk of hallucination: the model generating features, edges, or textures that look convincing but don't correspond to anything on the ground. For a general audience, a hallucinated rooftop texture might be harmless. For clients in mining, exploration, environmental monitoring, or government applications, where imagery may inform decisions about infrastructure, compliance, or land change, a fabricated detail is a serious problem if it isn't understood and managed.

This is the tension sitting at the heart of AI-driven super-resolution today: real gains in apparent clarity, offset by real questions about what you can and can't trust in the output, and under what conditions.

Join Us to Explore It Properly

We're hosting a webinar to unpack all of this in detail: where super-resolution technology currently stands, the practical benefits we've observed applying it across different sensors and use cases, and the challenges and limitations users need to understand before relying on it. We'll be drawing directly on our own testing and results from the past 18 months.

Register Here

Want to work with us? Let’s talk

We are the experts in satellite imagery and geospatial solutions