Blue tab
Green tab
Brown Tab

Preliminary results (examples)

 

 

Test Site 1. - Sokolov - Hyperspectral flight line leveling


Background

Endmember extraction and the preliminary thematic maps were produced using the reflectance data with no cross-track illumination correction, BRDF correction, or radiometric leveling across flight-lines. As such, DLR has been developing a novel method to address these problems. For hyperspectral surveys that include multiple adjacent flight-lines there is commonly a need to create a “seamless” mosaic, which is done for visual continuity, but also to remove line to line inconsistencies. Empirically-based approaches have been developed, where the common approach is to normalize the flight-lines to a “master” flight-line (e.g. Palubinskas et al., 2003). This approach suffers in that errors can propagate and reduce quality for flight-lines further from the “master” line. Overlap areas between adjacent flight-lines are commonly compared with respect to different land cover classes, with each having a different BRDF model. This approach requires pre-classification and knowledge of the different scattering properties of surface materials.


Methods

The new leveling approach being developed differs in that it does not use a “master” line, but instead uses all lines equally to derive a correction factor to level the flight-lines. In addition, we do not pre-classify an image into surface types and correct for each pixel based on the class type. Instead we generate a per-pixel scattering correction that is applied to the reflectance image prior to leveling. The basic concept for the scattering correction is that some surfaces, such as a concrete parking lot, should appear similar regardless of the geometry between sensor and object. On the other hand, densely vegetated areas can appear quite different due to a more pronounced BRDF and complex canopy shading effects. This is observed in predominant leafy forest cover where flight-lines are perpendicular to the sun. Thus, the relative degree of scattering can be related to the degree of vegetated cover. Using this assumption we can build a relative estimate of scattering on a pixel by pixel basis by using a relationship between per-pixel vegetation cover and albedo. Once the scattering correction has been applied a leveling process is applied to reduce line-to-line radiometric differences. This approach makes use of the overlap statistics between all pairs of flight-lines to derive a correction factor that is applied to each line and weighted based on a pixels relative location to nadir.


Results

The following figure shows the steps of the scattering correction applied to a portion of one of the Sokolov 2010 flight-lines. In figure part D forested areas show increased relative scattering towards the bottom of the image, specifically an increase in shadow within the canopy. Agricultural fields with variable vegetation cover show moderate relative scattering, whereas bare soil and urban materials show little or no scatter. Figure part E is the corrected reflectance image, which is used for leveling. In this image the forested areas appear “flattened” with shaded areas less pronounced. In addition, agricultural fields with different amounts of vegetation versus soil cover also appear to be more pronounced. The last figure shows the results from leveling with respect to RGB combinations that include the visible, near-IR, and SWIR regions. In each case the results show that the leveled product significantly reduces line to line inconsistencies. The combination of using a pixel by pixel relative scatter correction and flight-line leveling that does not correct to a single “master” line shows promising results for image mosaicking.

Click on image to see it full-sized.

HyMap flight-line images
Portion of one HyMap flight-line showing: (A) RGB true colour, (B) relative albedo, (C) vegetation index,
(D) per-pixel relative scattering correction, (E) corrected reflectance.
HyMap flight-line images
Leveling results for the HyMap flight-lines showing different RGB band combinations (bottom row). Unprocessed data is shown in the top row. Top left inset shows a portion of a coal mining region, whereas the bottom right inset shows an area dominated by forest and agricultural areas.