Change Detection

Change detection is what imagery becomes once you stop describing a place and start comparing it. Where the Time Series concept arranges repeated observations into a signal, change detection formalises the step from that signal to a labelled, validated statement that something changed — where, when, and by how much. The question is deceptively simple, but answering it well means being careful about what kind of change you mean, what you are comparing against, and how much of any apparent difference is real. These concerns belong to the observations and the reasoning about them, so they hold whatever software you later use to compute a difference or draw a change map.

What counts as change

“Change” is not one thing, and naming which kind you are after decides how you measure it. The simplest notion is a difference: subtract one date’s values from another’s and read where the result is large. A second is a trend — a gradual, directional drift estimated across many dates rather than two, the slow gain or loss that only a longer record reveals. A third is a threshold crossing: change is declared only where a difference exceeds some magnitude, separating meaningful departures from ordinary fluctuation. A fourth is a class transition, where each date is first labelled — forest, water, built, bare — and change is the move from one category to another between dates.

These map onto the standard families of method. Image differencing and index differencing compare continuous values date to date and are quick and transparent, but they answer “how much did the number change” rather than “what became what”. Post-classification comparison instead labels each date on its own, as the Classification Basics concept describes, and then compares the labels, so it names the transition directly and tolerates sensor differences between dates — at the cost of inheriting the error of two independent classifications. Choosing among them is really choosing which question you are asking, and each produces a different kind of answer that must be read on its own terms.

Comparing before and after

The most familiar design is bitemporal: one image before an event and one after, compared to isolate what the event did. Its strength is directness, but its validity rests entirely on the two dates being comparable in every respect except the change of interest. That usually means choosing images from the same season, or even near the same calendar date in different years, so that the surface is in a similar phenological and illumination state on both — an anniversary pairing — rather than comparing a summer scene against a winter one and calling the season a disturbance.

A two-date comparison also carries an ambiguity: it shows that the endpoints differ, but not the path between them, so a change and a later reversal can cancel and a transient event can be missed entirely. Continuous, time-series-based detection addresses this by modelling the whole record and flagging where a new observation departs from what the recent history predicts, which localises change in time and separates an abrupt break from a gradual drift. Before-and-after is the right frame for a known event at a known date; monitoring across a full series is the right frame when the timing itself is what you are trying to discover.

Separating meaningful change from normal variation

Not every difference between two dates is change worth reporting. Landscapes vary constantly for reasons that are entirely expected: vegetation greens and senesces through its seasonal cycle, water bodies rise and fall, tides and floods advance and retreat, and the same surface reflects differently as the sun angle shifts through the year. A difference image computed across these will light up everywhere, most of it describing the ordinary rhythm of the place rather than any event. The interpretive task is to hold this normal variation as the baseline and ask what departs from it.

This is why the temporal context developed in the Time Series concept matters so directly. A single dip means little until you know whether it falls within the usual seasonal trough; a rising value is recovery in one context and encroachment in another. Distinguishing signal from expected variation draws on domain knowledge of what the surface should be doing — the phenology of the land cover, the local climate, the management calendar — so that a threshold or a model is judged against a realistic expectation rather than against zero.

Consistency: everything equal but the ground

A detected change is only trustworthy if the numbers changed because the ground changed, not because the observation did. Several conditions must hold for the two dates to be genuinely comparable. The scenes must sit on a common, well-registered grid so that a given pixel follows the same patch of ground on both dates — the co-registration requirement from the Coordinate Systems concept — because a misalignment of even part of a pixel makes every edge and boundary read as change that never happened, a false signal that traces the outlines of features. Values should be calibrated to comparable physical units rather than raw counts, and ideally corrected for atmospheric effects, so that haze or a different aerosol load on one date is not mistaken for a surface change.

Cloud, shadow, and haze are the most common contaminants, and undetected cloud is a classic source of spurious change; screening them out first, as the Cloud Masking concept sets out, is a precondition for any honest comparison. Working in a normalised index rather than a single band helps further, since a ratio of the kind described under Spectral Bands largely cancels illumination differences that scale bands together and isolates the surface property of interest. The recurring principle is that every difference between two images except the one you care about is a potential false alarm, so the craft of change detection is as much about making everything else equal as about measuring the change itself.

Validation and uncertainty

A change map is a hypothesis, not ground truth. Once produced, it must be checked against independent reference information — field records, higher-resolution imagery, or expert interpretation of the scenes — at a sample of locations, so that its reliability can be stated rather than assumed. Two kinds of error are always in tension: omission, where real change is missed, and commission, where stable ground is wrongly flagged as changed. A threshold set to catch every event will raise many false alarms; one set to suppress false alarms will overlook subtle change. There is no setting that removes both, only a defensible balance for the decision the map serves.

Reporting that balance honestly is part of the result. A change product carries uncertainty from the sensor, the registration, the residual atmosphere, the threshold or model, and the reference data used to check it, and a stated accuracy — how much of the flagged change is real, and how much real change was caught — is what turns a coloured map into evidence a decision can rest on. Read with that uncertainty in view, together with the comparability that made the two dates fair to compare and the normal variation held as a baseline, change detection becomes what it is meant to be: a disciplined account of how a place has actually changed, and how confident we can be that it has.