For example, we classified the image for Path 28 Row 30 from August 26, 1991, which had limited reference
data (nine SDT measurements, all < 1 m, within seven days of the image date) based on the model developed from
the image for Path 28 Row 29 on the same date. This image had 25 SDT measurements within seven days of the
image date. A comparison of the classified lakes from the overlap area of both images indicates very good
agreement (R2 = 0.996 and SEE = 0.48 TSI(SDT) units). For better results we suggest using data from both or all
images in the path to increase the available number of observations and range of data used to calibrate the imagery.
We also conducted an experiment using an adjacent scene from a different path and time to calibrate a Landsat
image. The first image (Path 28 Row 28 (east) from August 7, 1990) had 140 SDT measurements within 7 days of
the overpass date. The regression model for ln(SDT) versus Landsat spectral-radiometric data for this scene has an
R2 = 0.81 and SEE = 0.31. The adjacent scene (Path 29 Row 28 (west) from August 17, 1991) had 57 SD
measurements within seven days of the Landsat overpass date. The regression model for ln(SDT) versus Landsat
spectral-radiometric data for this scene has an R2 = 0.81 and SEE = 0.29. For the experiment we used Landsatinferred
values of ln(SDT) from the “east” image as reference data for the overlapping lakes of the “west” image.
This resulted in 714 reference lakes and a regression model (inferred ln(SDT) versus Landsat spectral-radiometric
data) with an R2 of 0.58 and SEE = 0.50. To improve the relationship we limited the reference lakes to those for
which >100 pixels had been used to calculate the spectral mean. This reduced the number to 330 reference lakes and
yielded a regression model with R2 = 0.67 and SEE = 0.48. The latter relationship was used to calculate lake water
clarity for the “west” image. A statistical comparison of water clarity results from the SDT reference data model and
the adjacent scene reference data model for the “west” image indicate a very strong relationship with an R2 of 0.95
and SEE = 2.25 TSI(SDT) units.
Finally, we found that use of an image from the same path and row but different time period to calibrate other
images is feasible as long as some method of radiometric calibration is applied (Kloiber et al.,2002). However, the
resulting predictions from this method have a greater uncertainty than those obtained using in-situ data to develop a
scene-specific regression relationship. Therefore, it is preferable to develop scene-specific relationships with in-situ
data whenever possible. In the future, as better atmospheric correction technologies become available, other
approaches may prove to be as accurate.