70
DC restore (ADC) from the previous scan, and the average calculated before DC restore (BDC) from the current scan, a linear
interpolation was performed to determine and remove the bias at each pixel location on the current scan. The first scan of an
image was processed with the original technique and window w6 since no previous scan data was available.
Notethatthe beginning of each darkshutter window does not occur simultaneously for every detector. Because of the shutter
mechanism's configuration, detector one "sees" the dark shutter before detector sixteen in reverse scans. Inthat case window
placement must be shifted slightly in time for each detector to obtain the proper, corresponding shutter values. Table 1
presents the locations of the final extraction windows chosen for Landsat 5.
SHUTTER EXTRACTION WINDOWS
DI ANN BEFORE DC RESTORE START | AFTER DC RESTORE START NO. OF MINOR
MINOR FRAME MINOR FRAME FRAMES
Forward 109 538 52
Reverse 342 728 (Det. 1) - 781 (Det. 16) 52
Table 1. Final Landsat 5 TM shutter extraction parameters
6. RESULTS
The previous processing scheme was tested on the following Landsat 5 images.
WRS Path=188 Row=041 Libyan Desert collections:
8 September, 1984
29 August, 1986
1 September, 1987
9 September, 1990
27 August, 1991
WRS Path = 033 Row = 033 Colorado Springs: 27 July, 1985
WRS Path = 172 Row = 066 Zambia: 20 June, 1984
WRS Path = 172 Row = 071 Zambia: 20 June, 1984
WRS Path = 172 Row = 072 Zambia/Zimbabwe: 20 June, 1984
WRS Path = 172 Row = 076 Botswana: 20 June, 1984
Unfortunately, the pre-1991 Libyan Desert images provided shutter data that was pre-averaged over the window designated
in the Landsat 5 Long Term Parameter File, thus making a analysis of those images with the new window impossible.
However, it was still possible to evaluate the impact of the newly calculated gain coefficients on levels of streaking in all of the
images. The new detector gain coefficients calculated from the earlier1984 Zambian image can be found in Table 2. Figure 3
provides a sample comparison of the relative detector gains provided in the Long Term Parameter file with the image, to the
detector gains determined fromthe equalization procedure withthe 1991 Libyan desert scene and the 1984 scene of Zambia.
The 1984 and 1991 equalization values are virtually identical while the Long Term Parameter values are clearly different
graphically indicating the reason for the streaking in images produced with the Long Term Parameter gains and that the
relative gains of the detectors have been extremely stable over many years.
Streaking performance was consistently good across all years and for all of the reflective bands exceptband 5, notbecause of
an inherent deficiency in the process, but because data were unavailable for determining new coefficients for band 5 since it
saturated in the 1991 image. A quantitative determination of streaking (here defined as the difference between the average
value of one line from one detector and the average value from the two surrounding lines from two adjacent detectors) was
sometimes difficult to establish because of scene content. However, residual streaking values varied from a high of less than
0.5DNtoalowof 0.015 DN. Band 2 consistently provided the highest amount of residual streaking. More variation of residual
streaking was seen from band to band than from image to image.
Itwas also discovered during the study that residual streaking levels were so small that the quantization of the final pixel values
to 8 bits (256 gray levels) from the calculated real numbers was actually serving to increase visible streaking. An along track
profile through a lake in the Colorado Springs image is presented in Figure 4 quantized to 8 bits and 15 bits. The additional,
stairstep variations displayed in the 8 bit quantized profile clearly serve to exacerbate small differences and cause streaking
and striping.
Since striping compensation was calculated on a scan by scan basis using only dark shutter data, its magnitude is virtually
independent of the equalization process. With very little uniform image data accompanied by raw shutter data, residual
striping has also been very difficult to quantify. Using 15 bit quantization, the only way to effectively judge the degree of residual
striping is by visual softcopy inspection and a contrast stretch of from 20 to 80 times (i.e. mapping 8 DN into 256 gray levels).
The best estimates of performance with these procedures placed striping (which was defined simply as the average difference
between two adjacent scans) at an average of approximately 0.15 DN, with a maximum of about 0.5 DN. Striping levels after
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995
"ma m muda S J3db 65 dA
ff 9, 7 "yg FN — s Foo