Change detection compares Earth-observation data from different dates to identify what has changed, where and often when — deforestation, urban growth, new construction, flood impact, mining expansion. It sounds simple: subtract then from now. In practice, distinguishing real surface change from differences in season, illumination, atmosphere and sensor is the entire craft.
Core techniques
Bi-temporal methods compare two co-registered, comparably processed images: band or index differencing, ratioing, and post-classification comparison (classify each date, compare labels). Post-classification conveniently yields from-to change matrices, but multiplies the errors of both input maps.
The stronger modern paradigm is dense time-series analysis: algorithms in the spirit of LandTrendt, BFAST and CCDC model each pixel’s full trajectory across dozens or hundreds of observations, detecting breaks, trends and seasonality — and flagging change with a date attached.
The pitfalls that create false change
Misregistration is pitfall one: images offset by even a pixel generate rings of phantom change along every edge. Pitfall two is phenology — comparing March to September detects mostly the calendar. Pitfall three is radiometric inconsistency between dates or sensors, which mimics broad subtle change.
Disciplined workflows therefore fix registration to subpixel level, compare like season with like, correct to surface reflectance, and mask clouds and shadows aggressively — before any change algorithm runs.
From change map to monitoring service
One-off change maps answer audits; operational value comes from monitoring: scheduled acquisitions, automated processing, alerts pushed when change exceeds thresholds inside areas of interest — encroachment on a right-of-way, clearing inside a concession, new structures in a floodway.
The deliverable matures from a map to a feed: polygons of change with dates, confidence scores and imagery evidence, delivered into the client’s GIS or workflow system.