Change Detection

Change Detection maps what happened between two acquisitions: deforestation, urban growth, flooding, fire damage, crop rotation.

The five tools differ in what they compare, and the choice matters more than any parameter.

Before comparing anything

Change detection fails for the same three reasons every time, and all of them are avoidable.

The images must line up. A misregistration of one pixel produces a false edge of change along every boundary in the scene. Check the alignment before anything else.

The images must be radiometrically comparable. Two scenes taken months apart differ in sun angle, atmosphere and sensor calibration. Correct them both, in Radiometric Correction, or at least run Histogram Matching Two Images from Enhancement & Contrast. Otherwise you will map the weather.

The dates must be comparable. Comparing June with December on farmland maps the growing season, not land cover change. Use the same season, unless the seasonal difference is precisely your subject.

Start here

Image Difference Change Detection

The direct approach: subtract one date from the other, band by band, and map where the difference is large. Simple, transparent, and the right first step.

It tells you where something changed and how much, not what it changed into. That is often enough, for instance to direct attention to the areas worth examining.

Thresholding the difference is the whole difficulty. Small differences exist everywhere, from noise and from slight differences in conditions. Otsu Thresholding, in Classification, gives a defensible automatic threshold.

Change Vector Analysis

The refined version. Treating each pixel’s spectral values as a vector, it computes the magnitude of change, how much moved, and its direction, in which spectral sense it moved.

The direction is what makes it valuable: vegetation cleared and vegetation growing both produce a large magnitude, and only the direction separates them. Two rasters come out, and they are read together.

Post Classification Change

The most informative approach, and the most demanding. Classify each date independently, then compare the two class maps pixel by pixel. The result is a transition matrix: forest to bare soil, bare soil to built up, and so on.

The catch is well known. The errors of both classifications combine, so two maps at 90 percent accuracy give a change map at around 80 percent, and much of the apparent change is classification noise. Use it when you need to know what became what, and be strict about the accuracy of each date.

Pca Based Change Detection

Stacks both dates and runs a principal component analysis over them. The first components capture what the two dates share, and the later ones capture what differs, so change concentrates in those.

Effective when the change is subtle and hard to threshold directly, and it needs no training data. The components require interpretation, which is the usual price of the method.

All the tools

Tool	What it does
Image Difference Change Detection	Pixel by pixel difference between two coregistered multispectral dates.
Change Vector Analysis	Magnitude and direction of the spectral change between two dates.
Pca Based Change Detection	Principal components of the stacked dates, isolating change in the later components.
Post Classification Change	Compares two independently classified maps and quantifies the transitions.
Write Function Memory Insertion	Packs two or three dates into an RGB composite, so change appears as colour.

Notes

The colour composite trick. Write Function Memory Insertion produces no statistics, only an image, and it remains one of the fastest ways to read change: put each date in a different colour channel and anything that stayed the same comes out grey while anything that changed comes out coloured.

Radar sees through clouds. For flood mapping and for monitoring in the tropics, optical imagery is often unusable on the date that matters. See SAR.