Computer Vision on Satellite Imagery: Scale Is the Advantage, Not Perception

Computer vision applied to satellite imagery is the set of techniques that let a machine automatically identify, classify, and georeference events on the ground. The change this represents for industrial operations is, at its core, a change of scale: a human analyst takes an hour to review a high-resolution image, while a well-trained model processes the same area in seconds — which for mining, transmission, and oil and gas means continuous detection that doesn’t depend on how many analysts are available.

What the machine sees that the human eye can’t scale

Computer vision in remote sensing works on a combination of color, texture, shape, and temporal variation that a human analyst would identify without difficulty, but never at the speed or volume needed to cover an entire operation. A convolutional neural network learns from examples annotated by experts and, once trained, classifies new events in a fraction of a second — detecting encroachment on a right-of-way by geometry and estimated onset date, identifying vegetation clearing through temporal comparison between images, and finding fire hotspots by their thermal signature before smoke is even visible to the naked eye.

A frequently underestimated point is that a model calibrated by biome and asset type outperforms a generic model, because the spectral signature of the Cerrado differs from the Atlantic Forest, just as the texture of a PRAD area differs from an active operational perimeter. The model doesn’t tire, doesn’t lose focus, and processes everything at once. Scale, therefore, is the advantage, not perception itself.

How the pipeline turns pixels into operational alerts

Between the arrival of the orbital image and the generation of the alert, a sequence of automatic steps eliminates noise and increases the precision of what reaches the technical team. Pre-processing corrects for atmosphere, reprojects, and normalizes the image to the client’s coordinate system — a step without which comparing dates produces artifacts. Next, inference applies the neural network to every pixel, combining geometry, spectrum, spatial context, and temporal history in a single calculation, and the classified event goes through cross-validation against another data source, another sensor, another date, or the historical series, ensuring a false positive eliminated at this stage never becomes a notification. Only then comes contextualization by asset, where the same event gets a different interpretation depending on where it occurs, and finally the georeferenced alert — with polygon, coordinates, date, and asset context — ready to open a ticket without a prior confirmation visit.

Why per-asset calibration separates useful detection from noise generation

A generic computer-vision model, trained on global data, tends to produce an unacceptable false-positive rate for a regulated industrial operation, because every asset carries a pattern of expected change that needs to be discounted before any event becomes an alert. A waste pile growing according to the mining plan isn’t an event, but the same pile growing beyond its limit is — just as vegetation regrowth within a PRAD is successful recovery while the same regrowth outside it is unauthorized clearing. Calibrating the model to each asset’s specific rules is what determines whether the system gets used or ignored by the technical team, because a system that generates excess noise loses trust fast.

Where the human analyst remains irreplaceable

Human oversight remains necessary at three points in the cycle, and no model advance eliminates all three at once. The first is annotating training examples, since the network learns from what the specialist labels, and bad examples produce a bad model. The second is an event with no precedent in the historical series, which the model classifies with low confidence precisely because it hasn’t seen anything similar before, leaving the decision to the analyst. The third is the regulatory decision itself: the model detects, classifies, and prioritizes, but the signature on the administrative process always stays with the technical lead.

How NOR applies computer vision in real mining and transmission operations

At NOR Space Intelligence, every model is calibrated by asset and by biome before going into production, according to asset type, biome, and the company’s operational goal, which means no client starts with a generic model. At one mining company, the models distinguish over a thousand land assets by category — dams, PRAD areas, operational perimeters, and land under licensing — applying distinct monitoring criteria for each type. At a transmission company, the system automatically compares images captured before and after right-of-way clearing campaigns against the planned schedule, confirming execution within seven days, with a 94% detection rate of completed campaigns.

This entire model layer runs on infrastructure designed for traceability: from image receipt to the issuance of the georeferenced alert, every processing step is versioned, receives an integrity hash, and runs on national infrastructure, forming an evidence chain auditable by agencies like ANM, ANEEL, and IBAMA. For new clients, this translates into a demonstration run on the operation’s actual area even before a commercial proposal, showing detection happening on the client’s own asset, along with accuracy and cost-per-detected-event estimates.

Anyone who wants to see this process applied to their own operation can schedule a demo with NOR.