Change detection at 10 m resolution: Sentinel-2 workflows in production

Sentinel-2 is close to a gift: 10 m multispectral imagery, a five-day revisit, and no licence fee. For land-cover monitoring that is a compelling baseline. But "free imagery" and "reliable change detection" are very different things, and the gap between them is where most projects quietly fail. Naive differencing of two dates produces a change map that is mostly cloud edges, illumination differences, and atmospheric haze — real change buried in noise.

Clouds are the first and hardest problem

A single Sentinel-2 acquisition over a tropical region is often more cloud than ground. You cannot difference what you cannot see, and a cloud on one date reads as dramatic "change" against a clear pixel on another. The fix is not a better cloud mask alone — it is compositing: build a cloud-free mosaic from multiple acquisitions in a window, choosing the clearest pixel per location, so both of your comparison dates start from clean ground.

Radiometric normalisation before you difference anything

Even cloud-free, two dates are not directly comparable. Sun angle, viewing geometry (the bidirectional reflectance effect), and atmospheric scattering all shift pixel values without any real change on the ground. Differencing raw reflectance turns those shifts into false positives. Normalising the two dates to a common radiometric baseline — correcting for illumination and atmosphere first — is what makes the subtraction meaningful.

• Composite to cloud-free mosaics per comparison window.
• Apply atmospheric correction so you compare surface reflectance, not top-of-atmosphere.
• Normalise for sun angle and view geometry to suppress BRDF-driven false change.

Per-class confidence thresholds

A single change threshold across all land-cover classes is a blunt instrument. The spectral signature of genuine deforestation is loud; the signature of subtle cropland-to-fallow transition is quiet, and a threshold tuned for the former misses the latter while a threshold tuned for the latter drowns in noise. We set thresholds per class, calibrated against validated reference change, so each transition type is judged on its own signal.

Making it hold across scene types

The same pipeline runs over tropical and arid regions, which behave nothing alike — one fights cloud, the other fights bright soil and sparse vegetation. Keeping the false-alarm rate low across both came down to compositing and normalisation doing their jobs before detection, and to thresholds calibrated per scene type rather than assumed constant. The detector is almost the easy part once the inputs are clean.