Detecting change from space

Formal setup#

Given two co-registered images I_1 and I_2 over the same spatial extent, define a change map C:

D(x, y)  =  f( I_2(x, y), I_1(x, y) )

C(x, y)  =  1  if  D(x, y) > T
            0  otherwise

• I_1, I_2 — the earlier and later images, each a function from pixel location (x, y) to a value or vector of band values.
• f — a comparison function. Simple differencing uses f(a, b) = |a - b|. Band ratioing uses f(a, b) = a / b, which is less sensitive to multiplicative illumination change than a raw difference.
• T — a threshold, either fixed, or set statistically as some number of standard deviations above the mean of D over a known no-change region.

Method families#

Algebraic methods — image differencing, band ratioing, change vector analysis (CVA). CVA generalises differencing to multiple bands by treating each pixel's band values as a vector and measuring both the magnitude and direction of the vector's shift between dates; direction indicates the type of change (e.g. vegetation loss vs. new water), magnitude indicates how much.

Post-classification comparison — classify each date's image independently into land-cover classes (forest, water, built-up, bare soil), then compare the class labels. This avoids radiometric normalisation issues entirely, at the cost of compounding two classifiers' errors — an error rate of e per classification date yields a combined error rate approaching 2e - e^2 for the change map, ignoring correlation between the two classifiers' mistakes.

Deep learning approaches — Siamese convolutional or transformer encoders process I_1 and I_2 through shared weights, and a decoder head predicts the change map directly from the two feature representations. Trained end-to-end, these implicitly learn robustness to the confounds a hand-tuned threshold cannot handle, at the cost of needing labelled change/no-change training data, which is scarce and expensive relative to single-date land-cover labels.

Radiometric normalisation#

Before differencing, dates are commonly normalised to reduce illumination and atmospheric differences unrelated to real change. Pseudo-invariant feature (PIF) normalisation regresses one date's reflectance against the other over pixels assumed stable (roads, exposed rock), then applies the fitted linear transform to the whole image before comparing.

Evaluation#

Change detection is typically framed as pixel-wise binary classification, evaluated with precision, recall and the kappa coefficient against a human-labelled reference change map — see model evaluation for the general metrics. Kappa corrects for the class imbalance that dominates most real scenes, where the overwhelming majority of pixels do not change in any given interval and raw accuracy is misleadingly high by default.

Papers and current state#

• Chen, J. et al. (2020), DASNet: Dual Attentive Fully Convolutional Siamese Networks for Change Detection, a widely cited deep architecture for the paired-image setting.
• Daudt, R. C. et al. (2018), Fully Convolutional Siamese Networks for Change Detection, ICIP — an early and still commonly used baseline architecture.
• The LEVIR-CD and OSCD datasets are the standard public benchmarks for building and general change detection respectively.
• Global Forest Watch's deforestation alerts (Hansen et al., University of Maryland, published continuously) are a production system built on this family of methods at country scale, run against Landsat and Sentinel data.

What to learn next#

• Anomaly detection — thresholding and scoring unusual observations in general.
• Convolutional neural networks — the encoder architecture behind most learned change detectors.
• Model evaluation — precision, recall and the imbalance problem in depth.