Resampling
Where do resampling occurs?
When a satellite image is transformed geometrically, a new image based on a projected
coordinate system is formed. But the new image has no pixel values. The pixel values must
be filled through resampling. Resampling in this case means filling each pixel of the new
image with a value or a derived value from the original image.
Resampling is needed whenever there is a change of cell location or cell size between the
input raster and the output raster.
Examples are:
When projecting raster from one coordinate system to another coordinate system requires
resampling to fill in the cell values of the output raster.
When a raster changes from one cell size to another resampling is needed such as you go
from a 5-meter cell size to a 10-meter cell size, cell size will be different in the output raster
grid.
Input raster
Different
cell size
Resampling
Methods
Output raster
But it’s not always an easy choice which resampling method to use because there’s more than
one way to recalculate cell values. There are four common ways to resample raster grids in
GIS.
• Nearest Neighbor Resampling
• Bilinear Resampling
• Cubic Convolution Resampling
• Majority Resampling
Nearest Neighbor Resampling
The nearest neighbor technique assigns the value of the cell whose center is closest to the
center of the output cell.
For processing speed, it’s generally the fastest because of its simplicity.
Because no values are altered in the output raster data set, nearest neighbor resampling is
ideal for categorical, nominal and ordinal data.
In the given image, the output raster is
resampled from a rotated input raster. The cell
centers of the input raster are in gray. The value
for one of the cell on the output raster (in red) is
derived by identifying the nearest cell center on
the input raster (the blue spot) and assigning its
value to the output cell center.
When is nearest neighbor resampling used?
Discrete data like land cover classification, buildings and soil types have distinct boundaries
and their limits are considered discrete. When you resample this type of data, you should use
nearest neighbor resampling.
For example – if you have a land cover classification raster grid, nearest neighbor will take
the cell center value. If agriculture has the discrete value of 7, the nearest neighbor method
will never assign it a value of 7.2. It simply involves taking the output value from the nearest
input layer cell center.
Bilinear Interpolation
Bilinear interpolation uses the value of the four nearest input cell centers to determine the
value on the output raster.
The new value for the output cell is a weighted average of these four values, adjusted to
account for their distance from the center of the output cell.
Since the values for the output cells are weighted based on distance, and then averaged, the
bilinear interpolation is best used for data where the location from a known point or
phenomenon determines the value assigned to the cell.
Bilinear Interoplation
Bilinear Interpolation
It smoothens the output raster grid, but not as much as cubic convolution. It’s useful when
working with continuous data sets that don’t have distinct boundaries.
When is bilinear resampling used?
Temperature gradients raster, digital elevation models, slope, annual precipitation grids,
noise distance raster – these are all examples of when bilinear interpolation is used to
resample images. Notice how each of these types of data vary continuously cell-to-cell to
form a surface.
Cubic Convolution
Cubic convolution is a resampling technique similar to bilinear interpolation except that the
weighted average is calculated from the values of the 16 nearest input cell centers.
In the given image, the output raster is resampled
from a rotated input raster. The cell centers of the
input raster are in gray. The value for one of the cell
on the output raster (in red) is derived by identifying
the sixteen nearest cell centers on the input raster
(the four blue spots) and assigning the weighted
average of the sixteen values to the output cell.
Cubic Convolution
As a result, processing time tends to increase for this method.
Compared with bilinear interpolation, cubic convolution has a tendency to sharpen the edges
of the input data since more cells are involved in the calculation of the output value.
This method is generally used for continuous surfaces where much noise exists. Because it
takes more neighboring cells compared to bilinear resampling, it’s good for smoothing data
from the input raster grid.
Majority Resampling
While nearest neighbor resampling takes the cell center from the input raster data, the
majority algorithm is based on the most common values found within the filter window.
Similar to the nearest neighbor algorithm, this technique is commonly used for discrete data
like land cover classification and other types of raster grids with distinct boundaries.
When is majority resampling used?
It’s most often used for discrete data. For example, if the filter window finds 3 cells of
agriculture land cover and 2 cells of road, the output data set will be classified as agriculture.
This is because the agriculture land cover class is the most popular cell within the filter
window. When compared with nearest neighbor resampling, the resulting data set will often
be smoother.