It is sometimes necessary to classify the elements in a dataset (for example pixels in an image) into a grid of individual, non-overlapping tiles. For example when background sky gradients are present in an image, you can define a tile grid over the image. When the tile sizes are set properly, the background’s variation over each tile will be negligible, allowing you to measure (and subtract) it. In other cases (for example spatial domain convolution in Gnuastro, see Convolve), it might simply be for speed of processing: each tile can be processed independently on a separate CPU thread. In the arts and mathematics, this process is formally known as tessellation.
The size of the regular tiles (in units of data-elements, or pixels in an image) can be defined with the --tilesize option. It takes multiple numbers (separated by a comma) which will be the length along the respective dimension (in FORTRAN/FITS dimension order). Divisions are also acceptable, but must result in an integer. For example --tilesize=30,40 can be used for an image (a 2D dataset). The regular tile size along the first FITS axis (horizontal when viewed in SAO ds9) will be 30 pixels and along the second it will be 40 pixels. Ideally, --tilesize should be selected such that all tiles in the image have exactly the same size. In other words, that the dataset length in each dimension is divisible by the tile size in that dimension.
However, this is not always possible: the dataset can be any size and every pixel in it is valuable. In such cases, Gnuastro will look at the significance of the remainder length, if it is not significant (for example one or two pixels), then it will just increase the size of the first tile in the respective dimension and allow the rest of the tiles to have the required size. When the remainder is significant (for example one pixel less than the size along that dimension), the remainder will be added to one regular tile’s size and the large tile will be cut in half and put in the two ends of the grid/tessellation. In this way, all the tiles in the central regions of the dataset will have the regular tile sizes and the tiles on the edge will be slightly larger/smaller depending on the remainder significance. The fraction which defines the remainder significance along all dimensions can be set through --remainderfrac.
The best tile size is directly related to the spatial properties of the property you want to study (for example, gradient on the image). In practice we assume that the gradient is not present over each tile. So if there is a strong gradient (for example in long wavelength ground based images) or the image is of a crowded area where there isn’t too much blank area, you have to choose a smaller tile size. A larger mesh will give more pixels and and so the scatter in the results will be less (better statistics).
For raw image processing, a single tessellation/grid is not sufficient. Raw images are the unprocessed outputs of the camera detectors. Modern detectors usually have multiple readout channels each with its own amplifier. For example the Hubble Space Telescope Advanced Camera for Surveys (ACS) has four amplifiers over its full detector area dividing the square field of view to four smaller squares. Ground based image detectors are not exempt, for example each CCD of Subaru Telescope’s Hyper Suprime-Cam camera (which has 104 CCDs) has four amplifiers, but they have the same height of the CCD and divide the width by four parts.
The bias current on each amplifier is different, and initial bias subtraction is not perfect. So even after subtracting the measured bias current, you can usually still identify the boundaries of different amplifiers by eye. See Figure 11(a) in Akhlaghi and Ichikawa (2015) for an example. This results in the final reduced data to have non-uniform amplifier-shaped regions with higher or lower background flux values. Such systematic biases will then propagate to all subsequent measurements we do on the data (for example photometry and subsequent stellar mass and star formation rate measurements in the case of galaxies).
Therefore an accurate analysis requires a two layer tessellation: the top layer contains larger tiles, each covering one amplifier channel. For clarity we’ll call these larger tiles “channels”. The number of channels along each dimension is defined through the --numchannels. Each channel is then covered by its own individual smaller tessellation (with tile sizes determined by the --tilesize option). This will allow independent analysis of two adjacent pixels from different channels if necessary. If the image is processed or the detector only has one amplifier, you can set the number of channels in both dimension to 1.
The final tessellation can be inspected on the image with the --checktiles option that is available to all programs which use tessellation for localized operations. When this option is called, a FITS file with a _tiled.fits suffix will be created along with the outputs, see Automatic output. Each pixel in this image has the number of the tile that covers it. If the number of channels in any dimension are larger than unity, you will notice that the tile IDs are defined such that the first channels is covered first, then the second and so on. For the full list of processing-related common options (including tessellation options), please see Processing options.