Radiometric Correction

A raw satellite image does not hold reflectance values. It holds digital numbers, arbitrary integers produced by the sensor, which depend on the calibration of the instrument, the position of the sun, the state of the atmosphere and the slope of the ground.

Radiometric Correction removes those dependencies, so that a value means the same thing across an image, across dates and across sensors. That is what makes an index comparable, a classifier transferable and a change detection meaningful.

Skip this sub toolbox only when your product is already delivered as surface reflectance, which most modern archives now offer.

The order of corrections

1. Digital numbers to top of atmosphere reflectance. Applies the sensor’s calibration coefficients and the sun geometry.
2. Atmospheric correction. Removes the haze the atmosphere adds, mostly in the blue and green.
3. Terrain correction. Compensates for slopes facing towards or away from the sun, which is essential in mountains and pointless on a plain.

Start here

Dn To Toa Reflectance

The first step. It converts digital numbers into top of atmosphere reflectance, using the gain and offset published with the scene and the solar geometry of the acquisition.

The values that come out are physical, between 0 and 1, and comparable between dates and between sensors. Everything downstream assumes this has been done.

Dark Object Subtraction

The simplest effective atmospheric correction, and the one to use when you have no atmospheric measurements, which is almost always.

Its principle is a good idea. Somewhere in the scene there is something genuinely black, a deep water body or a cast shadow. Whatever brightness it shows must have been added by the atmosphere. Subtracting that offset from the band removes the haze.

It works because it needs nothing but the image itself. It is approximate, since it assumes the haze is uniform across the scene, which it is not.

Terrain Corrected Optical Analytics

The full topographic correction, using a coregistered elevation model. It computes the illumination geometry from slope and aspect, corrects the reflectance accordingly, and returns the corrected stack together with a correction factor, a cloud and shadow mask and a confidence layer.

Without it, in mountainous terrain, a classifier learns the illumination rather than the land cover: the same forest on a sunlit and a shaded slope produces two different classes.

Brdf Normalization addresses the related effect where reflectance varies with the viewing and illumination angles, which matters for wide swath sensors and for mosaicking scenes acquired from different angles.

Correct Vignetting

For aerial and drone imagery rather than satellite. Lenses collect less light at the edges of the frame than at the centre, which shows up as a darkening at the corners and, once mosaicked, as a visible grid of darker seams. This tool removes it.

All the tools

Tool	What it does
Dn To Toa Reflectance	Converts digital numbers into top of atmosphere reflectance.
Dark Object Subtraction	Removes atmospheric haze using the darkest object in the scene.
Terrain Corrected Optical Analytics	Topographic correction using a DEM, with masks and a confidence layer.
Brdf Normalization	Normalises for viewing and illumination geometry, by C correction or Minnaert.
Correct Vignetting	Removes the radial darkening caused by lens optics.

Notes

You need the metadata. Converting to reflectance requires the calibration coefficients and the sun elevation, which come in the file distributed with the scene. Keep it alongside the imagery.

Correct before mosaicking. Correcting each scene first, then assembling, gives a far more coherent result than the reverse.

Then compute indices. NDVI and its relatives are meaningful on reflectance and misleading on raw digital numbers. See Normalized Difference Index in Enhancement & Contrast.