MapInfo Pro

Raster Resampling

The process by which a new pixel values are interpolated using existing pixel values whenever there is transformation of raster structure such as projection, reprojection changes, datum transformation, modifications through cell resizing is known as image resampling (Wade and Sommer, 2006).

Various resampling methods exists for resizing raster images i.e. whether image is enlarged or reduced, in which raster bounds remains unaltered and only transformation occurs in its structure. But it’s always a difficult choice that which resampling method to use as there are more than several ways to recalculate cell values.

Most popularly used resampling methods are nearest neighbor, bilinear and bicubic besides aggregated average, pixel resize and weighted average methods of resampling. 

Three most widely used resampling technique have been discussed below to provide insights on which one to use under given scenarios.

• Nearest Neighbor
• Bilinear
• Bicubic

Nearest Neighbor Resampling

The resampling method in which a pixel from the original image is matched to its corresponding position in the resized image is known as nearest neighbor resampling (Figure 1). If corresponding pixel is not available, the nearest pixel value is used instead.

Figure 1: Schematic representation of Nearest neighbor resampling method.

The grid represents the output raster. The center of the cells of input raster are marked with gray circles. The blue circle marks the center of the target cell. The yellow circle is nearest cells from the input raster that is used for calculating output raster cell value [green cell].

Major advantage of nearest neighbor resampling technique is that it preserves original values and no interpolated values are created and hence is closest representation of the image. In addition, because of the speed of implementation and simplicity it remains the widely used for image resampling. A notable disadvantage is that this method introduces position errors where the realignment of pixels is obvious especially along linear structures.

When to use Nearest Neighbor Resampling Method?

This resampling method is ideal for representing categorical data such as nominal, and ordinal data.

Nominal Data: Named/ Class data which can be separated into discrete categories

(Example - Soil Type: "Sand", "Silt", "Clay"; Land-Use Class:"Forest", "Open", "Urban")

Ordinal Data: Data placed into some kind of order or scale

(Example - Likelihood and incidence of risk - "Low" "Moderate" "High"; Classes numbered and named: "1 - Good," "2 - Moderate," "3 - Poor")

Bilinear Resampling

Bilinear resampling technique employs arithmetic mean of the nearest four (4) pixels of the target cell to calculate a new pixel value (Figure 2). Since, it’s an interpolated value so there is possibility of introduction of new value which is never present in original raster imagery.

Figure 2: Schematic representation of Bilinear resampling method.

The grid represents the output raster. The center of the cells of input raster are marked with gray circles. The blue circle marks the center of the target cell. The yellow circle are the four nearest cells from the input raster that are used for calculating output raster cell value [green cell].

When to use Bilinear Resampling Method?

Bilinear resampling methods provide smoother look and retains better positional accuracy than nearest neighbor resampling method and hence is useful for continuous data without distinct boundaries. But averaging sometimes brings local peaks and valleys and hence propagation of error is increased as compared to nearest neighbor.

Continuous Data: A surface in which each location has a value. 

(Example - Temperature, Precipitation, Elevation)

Bicubic Resampling

Resampling method in which a weighted average of the sixteen (16) pixels which are nearest to the focal cell is used is known as cubic resampling (Figure 3). Bicubic resampling is a variant of cubic resampling technique where the resampling is performed in both X and Y directions.

When to use Bicubic Resampling Method?

This method is ideal for continuous data where much noise exists. Resampling produces smoother surface as compared to nearest neighbor and bilinear resampling but is more computation intensive than other resampling methods as it takes more neighboring cells.

Why Raster Resampling in Important?

Geospatial Raster data viz. Satellite Imagery, Ariel Imageries, LIDAR, have different resolutions and coordinate system because of varied methods of data capture, provider preferences, satellite's inherent properties etc. In order to compare and analyze data from one spatial resolution/co-ordinate system to another, resampling techniques are being used such as nearest neighbor, bilinear interpolation, cubic convolution and majority interpolation. These various method of resampling have substantial effects on spatial properties desired for analytics. 

Nearest neighbor Resampling method is useful for its speed and categorical data integrity maintenance whereas Bilinear and Bicubic Resampling provide smoothness and accurately preserve mean values across many levels of aggregation of imagery. 

Thanks for posting Neena. I would like to add one correction and one addition.

Firstly, when we perform bilinear interpolation we interpolate between the cell centres of four adjacent cells arranged in a square. So the diagram is not quite giving the correct picture.

Secondly, we actually use two different methods of cubic interpolation, and the user is able to select which one they want. On the Raster ribbon, these two techniques are listed as "Cubic spline (local)" and "Cubic spline (global)" in the "Raster Quality" drop list. The local method is as you described above - it uses a 4x4 window of cells to make interpolations within the cell centres of the central four cells. The global method uses bi-directional cubic splines. We generate cubic splines along the rows and columns of the source raster level data, then interpolate data on those splines at the destination row and column positions. The interpolated data points are cubic splined again and then all cell values are interpolated from those splines.

The global technique produces smoother interpolated data and is generally preferred. However, all spline techniques produce undershoot and overshoot and, in some scenarios, this can cause extended ringing in the splines. In the second method, this ringing can extend throughout an entire data tile (which is the resolution at which the splines are computed). But if you use the local technique the ringing is quickly contained.

Which method you choose depends in part on how smooth your source (or base level) data is. If it is smooth then cubic splining will work very well. But if the data is not smooth you may wish to progressively dial back the interpolation method until you obtain an acceptable result.

Note that you are unlilkely to notice any significant performance difference between all these interpolation techniques when rendering a raster in Pro.

Thanks Sam for sharing detailed information here. Loved it!

The information shared on sampling is conceptual and not specific to Pro and hence there are difference in figure and content written and topic falls under GIS Essentials :)