4 Flat field correction

The flat field is separated in two in the calibration library: the optical flat and the detector flat. The goal of this separation is to minimize the number of configurations to calibrate. The data must be divided by the two flats. However, the library flat fields rarely correspond exactly to the observer's flat field as there exists a small jitter in the position of the wheels of the instrument. This jitter is intentional and was introduced to prevent the wheels from getting stuck. This means that for two observations with the same configuration, the optical flat field can be significantly different. Different methods have been tested in order to derive a flat from the data, but they apply only in the case of a raster observation.

In some cases, particularly for observations using a raster mode on faint sources embedded in a flat extended emission, e.g the zodiacal light, a flat field can be derived from the data by choosing interactively a region where there is no object, and only background. For an automatic flat field estimation, we take advantage of the fact that, in such observations, all pixels of the detector spend more time observing the background than the object, and we apply the following procedure:

1.

Average all frames belonging to the same configuration. This gives one image per configuration, i.e. a cube of averaged data $C_{\rm a}(x,y,c)$.

2.

For each pixel (x,y) of the detector, take the median of $C_{\rm a}(x,y,*)$. This results in M(x,y).

3.

The flat is equal to M normalized by the mean of its center: flat = M/mean(M(10:21,10:21)). This is the normalization used in all automatic and calibration data reduction, it therefore has to be employed in order to keep the data compatible with, for example, the ADU to mJy conversion factor.

The automatic flat field method generally produces good results. Its only limitation is the assumption that all pixels of the detector spend more time observing the background than the object. It will certainly not always be true for all observations, even in raster mode, and then the procedure will fail.

For raster observation, the flat field estimation can be improved by an iterative method, using the fact that a sky point is seen by different pixels (Starck et al. 1996; van Buren 1996).