|
ACRES General
Landsat
ALOS Related links 
Some of the documents on this webpage may be in PDF Format. Please download the Free PDF reader from Adobe to view these files.
|
Updated:
17 November 2006
Geometric correction and resampling of ACRES image dataRaw level satellite image data contains a number of geometric and geographic locational errors. Image geometric and locational errors arise from a number of sources. These sources include:
The geometric correction of satellite imagery involves the mapping of the original satellite data to a geometrically correct output grid. There are two major techniques used to correct the geometric distortion present in digital image data. One is to model the nature and magnitude of the distortions and locational errors in the uncorrected data and develop correction formulae. This approach is often referred to as the spacecraft model approach. The second technique is based on the establishment of a mathematical relationship between the addresses of pixels in an uncorrected image and their true position on the ground. This approach is based on the availability of suitable Ground Control Points (GCPs). The GCPs will most often be identified from a topographic map, GPS survey, or rectified image. The spacecraft model technique is used at ACRES in the generation of all ACRES products. For Landsat products, Control Points (CPs) are used in the refinement of the spacecraft model as not all sources of geometric errors in an image can be accurately modelled or measured. The CPs for most of ACRES orthocorrected products are geocoded Landsat ETM+ Panchromatic chips derived from accurately controlled full length Landsat 7 passes of the Australian continent. As part of the image correction process, brightness values must be derived for each cell in the image output grid. The value is interpolated from the uncorrected satellite image. This interpolation exercise is known as resampling. The spacing of that grid is chosen according to the pixel size required in the corrected image. The pixel size is usually similar to or slightly smaller than the spatial resolution of the geometrically uncorrected data.
Resampling kernels typically have two properties of interest, their mathematical foundations, and their length (the number of points over which the interpolation is performed). A kernel of length is called an n-point kernel. There are two classes of interpolator:
The former is simply standard polynomial fits of adjacent data points. The latter are derived from signal processing theory which states that, if a signal is band limited, it can be exactly reconstructed using a sinc (= sin (x)/x) interpolator. The ideal sinc kernel is infinite, however in practice it must be truncated to a finite length. This introduces truncation errors in the interpolator and, as expected, the shorter the kernel, the higher the truncation error. Available resampling kernelsIn the generation of ACRES products using data from optical sensors, the following resampling kernels are available
The choice of available resampling kernels for your product will depend on your intended use of the data. For Landsat spectral analysis, KD16 is the recommended option. Nearest Neighbour (NN)Performs no interpolation - rather the nearest pixel is selected. NN resampling does not alter the brightness values of the original image. It is often said that this resampling best reflects the original pixel values as observed by the sensor. However, NN resampling is not as visually appealing to the eye as other kernels. Cubic Convolution (CC)This is a 4-point kernel based on cubic splines. It has historically been used for remote sensing, as it is a reasonable compromise between accuracy and speed (the longer the kernel, the better the interpolation, but the slower the resampling speed). The CC kernel has a slight edge enhancing effect on the images. CC is the default resampling kernel for ALOS products. The high processing capability of the computers used for ACRES product generation results in little difference in the time taken for the generation of Landsat products using either the CC, KD16 or MTF resampling algorithms. The interpolation errors of the CC kernel are significantly worse than a 16-point kernel (below). Bilinear (BL)This is a simple 2-point linear interpolator that uses the neighbouring two points to produce a smoothing effect to the image. It is not used in most applications. 8-point Damped Sinc (DS8)This is an 8-point kernel based on windowing an 8-point truncated sinc kernel. Because of its increased length, it is a more accurate interpolator than the CC kernel. 16-point Damped Sinc (DS16)This is a 16-point kernel based on windowing an 8-point truncated sinc kernel. Again, it is slightly more accurate than the DS8 kernel. 16-point Kaiser-Damped Sinc (KD16)This is the most accurate kernel of the options presented here. It is based on a 16-point sinc function windowed by a Kaiser window. Internal studies have shown that it produces a pleasing balance between image ripple and low frequency accuracy, and has good RMS error results. It provides a more accurate representation than 4-point kernels such as CC. Modulation Transfer Function (MTF)The MTF resampling kernel is available only for Landsat 7 products. The MTF resampling kernel is based on an empirical modelling of the optical and electronic properties of the ETM+ sensor. MTF resampling kernel is only recommended for map and ortho corrected images. The MTF kernel may introduce a slightly blocky appearance to the more homogeneous areas of Landsat imagery. Additional details on commonly used ACRES resampling optionsOne of the following three resampling kernels are usually recommended for use in the generation of ACRES Landsat products:
The images below show a portion of a scene to better illustrate the difference in resampling kernels. To view the full image area of these examples, sample files are available for download. Nearest Neighbour (NN)
16-point Kaiser-Damped Sinc (KD16)
Modulated Transfer Function (MTF)
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||