Skip to main content
Methodology maintained by Elie Habib, founder of World Monitor. Published revisions are recorded in the corrections log.

Start here

Every few days a Sentinel-1 radar satellite passes over each watched land site, day or night and through cloud. WorldMonitor reads summary statistics of each new radar scene over a small box around the site and compares them with earlier scenes taken from exactly the same orbit. When a scene differs from that baseline by more than its normal spread and by more than a fixed floor, the site’s flag is raised. The next scene from the same orbit then either confirms the change or retracts it. A confirmed flag says only that imagery statistics changed. It does not say what changed or why. Flooding, harvest, snow, a drained reservoir, new construction and demolition can all move radar statistics. Read the flag as a prompt to look at the scenes, not as a finding. The flag is in a shadow period. During it, flags are computed and stored for review but are not shown on site evidence, country summaries or the API.

What is measured

A scene with less than 90% valid pixels over the footprint is no observation. A band whose statistic is missing is skipped, never read as zero. Neither moves the flag.

Comparable scenes

Radar backscatter depends strongly on viewing geometry, so a scene is compared only with scenes that share all of:
  • the constellation (Sentinel-1; its spacecraft share one baseline)
  • the relative orbit
  • the pass direction (ascending or descending)
  • the acquisition mode and polarization set
A change of acquisition mode starts a new baseline. A scene from another orbit can neither confirm nor retract a flag raised on this one. A site with no scene on an orbit for 180 days drops that orbit.

Baseline and change test

For each band statistic the baseline holds the newest 12 comparable scenes. It needs 6 before any scene is scored; until then the site is building its baseline. The centre is the median and the spread is a robust standard deviation (1.4826 × median absolute deviation), never less than 0.25 dB. A scene changes a statistic when its difference from the median is at least 3 spreads and at least 1.5 dB. A site’s state is the most significant state of its orbits: confirmed, then raised, then monitoring, then building its baseline. When a site is first scored, its baselines are filled from up to 6 archived scenes per orbit from the previous 120 days, so it starts monitoring without waiting for new passes. Each daily run spends a bounded share of its time on statistics requests; a site it does not reach keeps its flag and is scored on a later run from the scenes still in the 30-day archive window.

Shadow period and thresholds

The thresholds above were set from the spread of real scene statistics, and they will be revised from the shadow period’s distribution. The shadow period lasts at least 30 days and until at least 30 raised flags have been confirmed or retracted. Before flags are shown, a replay of 90 days over 30 sites is reviewed by hand and each confirmed flag is labelled as environmental or not. Observed during development, over 90 days ending 13 November 2026 on a spread sample of watched land sites, a typical comparable scene differed from its baseline median by a few tenths of a decibel. Coastal and river sites showed much wider spread from water surface roughness, and most of their raised flags were retracted by the next scene.

Limits

  • The statistic is a summary of a 2 km box. A change confined to a few buildings can leave the median and the 98th percentile unchanged.
  • Seasonal change (crops, snow, soil moisture, water level) is the most common cause of a confirmed flag. The flag does not attempt to remove it.
  • Sentinel-1 does not image every site on every orbit. Some sites have one orbit only, and a few have none in the archive.

Sources