TerrAlert Team · 2026-09-09
A satellite image, by itself, is just a grid of colored pixels. Turning that into something a person can use to understand what is happening on the ground takes several deliberate steps, and skipping any of them is where a lot of "satellite monitoring" products fall short of being genuinely useful.
Raw satellite data is affected by things that have nothing to do with the ground itself: atmospheric haze, the angle of the sun, sensor calibration drift between passes. "Level 2A" processing, the kind Sentinel-2 imagery is typically distributed as, corrects for atmospheric effects so that the colors and brightness in the image better reflect what is actually on the ground rather than what the atmosphere did to the light on its way up. Skipping this step means comparing two images that differ partly because of weather conditions on the day, not just because the ground changed.
A corrected image still needs to be placed precisely onto a coordinate grid, a process called georeferencing, so that "this pixel" reliably means "this exact spot on Earth" across every image, taken from every pass, regardless of the satellite's exact position that day. Without accurate georeferencing, comparing two images of "the same place" can actually compare two slightly different places, producing false differences that have nothing to do with real change.
Once two images are both corrected and aligned, they can be meaningfully compared, pixel by pixel or region by region, to find what is different. This is where a change-detection model does its work: looking at a "before" image and an "after" image of the same aligned area and identifying which parts differ enough to be worth flagging, versus which differences are just noise (a slightly different sun angle, a passing shadow, sensor variation).
A raw list of "these pixels changed" is not yet useful to a person. The last step is converting that pixel-level result into a clean shape, a polygon outlining the changed area, along with a rough size estimate and a category label (construction, forest loss, flooding, and so on). This is the difference between a scientific output and a product: a shape on a map with a plain-language label that someone can look at and immediately understand, versus a technical grid of probability scores that needs an expert to interpret.
A shape on a map means little without context. "0.4 km² changed here" is far more useful paired with a boundary showing whose land it is, what it was zoned for, and what it looked like a year ago than as a bare number. This is why monitoring tools are usually built around a defined Area of Interest rather than raw, unbounded imagery. The boundary itself is what turns a satellite observation into a specific, actionable piece of information about a place someone actually cares about. For the basic building blocks behind that boundary, coordinates, polygons, layers, see What Is Geospatial Data? A Simple Guide.
Why does the imagery need "correcting" at all? Because clouds, haze and sun angle change what a satellite records even when the ground itself has not changed. Correction removes as much of that noise as possible.
What does "Level 2A" mean? It is the processing level Sentinel-2 imagery reaches after atmospheric correction: surface reflectance, as close to "what the ground actually looks like" as the raw sensor data can get.
TerrAlert's own processing pipeline follows exactly these steps. The FAQ has the practical detail.
All posts