Skip to main content

Data quality and review confidence

Remote-sensing workflows are only as reliable as the inputs and context behind them. DeForsight makes it possible to record uncertainty instead of hiding it.

Check the input before the conclusion​

Before treating an alert as actionable, check:

  • Is the alert date inside the expected monitoring window?
  • Is the supplier or area boundary current and correctly located?
  • Is imagery available for the before-and-after comparison?
  • Are cloud, haze, shadow, or missing tiles limiting interpretation?
  • Does the change align with the alert source and confidence?

Common quality problems​

Stale geometry​

A supplier boundary may have changed while the record stayed the same. Confirm the geometry before treating an overlap as meaningful.

Cloud or haze​

Poor visibility can make a true change look weaker or create a false visual impression. Record the limitation and request a later acquisition when appropriate.

Timing mismatch​

An alert and an image can describe different dates. Always compare acquisition windows, not just the date on the record.

Duplicate or nearby signals​

Several alerts can represent one event. Use location, timing, and the review history to avoid counting the same event multiple times.

Confidence language​

Use precise language in notes:

  • Observed: the evidence directly shows the condition.
  • Supported: multiple pieces of evidence point in the same direction.
  • Inconclusive: evidence is incomplete or conflicting.
  • Not relevant: the signal does not relate to the supplier or monitored area.