Separate detail from map accuracy

A small pixel spacing describes the sampling grid, not the certainty of every coordinate. An image can show a crisp road while placing it tens of metres from its surveyed position. Conversely, a coarser image can be registered consistently enough for a broad-area comparison. Ask separately how much structure is represented and how accurately it is located.

ASAL’s early Alsat-2B commissioning account reports preliminary geolocation of 77 metres CE90 without ground control, against an initial 300-metre specification. Keep this historical result separate from the 2.5-metre panchromatic sampling figure. CE90 expresses a 90th-percentile horizontal position-error radius under the evaluated conditions, not the width of each pixel or an assurance that all errors are smaller.

Detail and position are separate

ASAL describes tests of dark response, detector uniformity, image sharpness and multispectral band alignment. These address different questions. A detector offset affects recorded brightness; a band displacement can create coloured fringes along an otherwise unchanged boundary. Neither should automatically be interpreted as a real difference on the ground.

In a fictional sensor model, measured value = gain × signal + offset. If the measured value is 110, the offset 10 and the gain 2, the corrected signal is (110 − 10)/2 = 50 arbitrary units. This teaching model is not an Alsat calibration equation or flight coefficient. It shows why subtracting an offset and correcting gain are different operations; neither operation establishes map coordinates.

Independent checkpoints help evaluate positional error. Points used to fit an adjustment should not be the only points used to claim its accuracy: that would test the fit against its own inputs. Record reference quality, the sample’s distribution, terrain and processing assumptions. A single well-aligned landmark is not a complete accuracy assessment.

Suppose a fictional dataset has a CE90 of 20 metres. The bounded interpretation is about its horizontal-error distribution under the test conditions. It does not promise that each future set of ten points will contain exactly nine inside a 20-metre circle, nor does it establish building-boundary accuracy for every location. Choose a product whose measured uncertainty fits the task.