Change Detection — Remote Sensing

Two images — one from a baseline date (t1) and one from the target date (t2) — are the minimum requirement. Having only two dates limits analysis to binary changed/unchanged classification. Multitemporal stacks of three or more images enable trajectory analysis, seasonal normalization, and more robust detection of gradual or episodic change patterns.

The choice depends on data availability and the type of change being monitored. Image differencing and NDVI differencing are straightforward when both images share similar sensor characteristics. Post-classification comparison is preferable when detailed from–to change matrices are required, though it compounds classification errors. PCA-based methods work well for exploratory detection when change direction is unknown. Singh (1989) provides a systematic comparison of these trade-offs.

Accuracy is reported using an error matrix derived by comparing the detected change map against independent reference points. Key metrics include overall accuracy, producer's accuracy (omission error), user's accuracy (commission error), and the Kappa coefficient. Reference samples should be stratified to ensure adequate representation of rare change classes, which are frequently undersampled in simple random designs.