Pan-sharpening algorithm causes unbalanced-color in overall image (GeoEye-1)

Discussion created by hlzhang525 on Feb 19, 2011
In our practices with raw bundle of GeoEye-1, we found that the pan-sharpening transformation method may bring too much histogram variance, which causes significantly-unbalanced color in overall image among imagery scenes in mosaic dataset (natural color band 321 with near-infrared weight value added).

So far, we have tested 3 isolated areas, i.e., 30-40 scenes/per area, totally, 125 scenes (roughly, 11,000 sqr km), within either one single mosaic dataset or seperate three datasets (http://help.arcgis.com/en/arcgisdesktop/10.0/help/index.html#//0017000000sw000000.htm and http://help.arcgis.com/en/arcgisdesktop/10.0/help/index.html#//009t00000041000000.htm). All results are not accepted by our customers.

Do we have any other approaches to apply for GeoEye-1 in order to get pan-sharpening transformation done and also can minimize histogram variance for better color-balanced image in overall?

Or, on color-balanced aspect, does anyone else have good practice with plenty of images in 'production' test and get acceptable results to share ?


In ESRI Image Server 9.3 /10 (image service definition file or mosaic dataset), ESRI algorithm for pan-sharpening is described below: (http://webhelp.esri.com/arcgisdesktop/9.3/index.cfm?id=3263&pid=3249&topicname=Using_the_Pan-sharpen_process ):

The ESRI pan-sharpening transformation uses weighted averaging (WA) and the additional near-infrared band (optional) to create its pan-sharpened output bands. The weighted average is calculated by using the following formula:
WA = (R * RW + G * GW + B * BW + I * IW) / (RW + GW + BW + IW)

The result of the weighted average is used to create an adjustment value (ADJ), which is then used in calculating the output values, as shown in the following example:
ADJ = pan image - WA
Red_out = R + ADJ
Green_out = G + ADJ
Blue_out = B + ADJ
Near_Infrared_out = I + ADJ

For the ESRI pan-sharpening transformation, the weight values of 0.166, 0.167, 0.167, 0.5 (R, G, B, I) provide good results when using QuickBird imagery.

It has been found that by changing the near-infrared weight value, the green output can be made more or less vibrant.