Test a hazard map against reference evidence

ASAL connects Alsat-3 with managing major risks. Turning imagery into a useful hazard layer nevertheless requires interpretation. A dark patch might be water, shadow or another surface; a missing view beneath cloud is unknown, not automatically safe. A mapping team needs a defined target class, observation times and suitable reference evidence.

Suppose a fictional team flags flooded locations from imagery and compares its decisions with a separate, time-matched reference survey. The exercise can reveal both false alarms and missed floods. If the reference describes yesterday while the image describes today, real change may complicate that comparison. Even the reference has limitations that should be recorded.

In our invented sample, the map flags twenty locations as flooded. The reference confirms flooding at fifteen of them, leaving five false alarms. Among the flagged locations, fifteen out of twenty are confirmed: 75 percent. This asks how reliable a positive map label is in the stated sample.

The reference survey finds twenty-five flooded locations in total, including ten that the map missed. The map therefore finds fifteen out of twenty-five reference floods: 60 percent. This asks how much of the target class was found. The two fractions have different denominators and answer different questions. In mapping terminology, these correspond to user’s and producer’s accuracy for the class in this simple exercise.

Two accuracy questions, two denominators

A false alarm could send a survey team to an unaffected location; a missed flood could leave an affected location without attention. The consequences depend on the decision, so a single headline percentage can be misleading. The sample counts are not a nationally representative estimate unless the sampling design justifies that inference.

A useful map note reports the target class, time window, reference method, confirmed and incorrect labels, missed cases and areas with no usable observation. A second sensor or local observation may resolve some unknowns. Do not erase an unknown region by coloring it as a negative result, and do not label the fictional percentages as measured Alsat-3 performance.