For this large regional assessment 19 Landsat images from five Landsat paths were needed to cover the entire
State of Minnesota. To date we have targeted two time periods ~1990 and ~2000 for statewide assessment. The best
results were obtained when entire Landsat paths of clear imagery were available for the same day. By using paths of
Landsat imagery of two to five Landsat images we are able to increase the available number of observations and
range of water clarity (because water clarity in Minnesota generally increases from south to north) in the dataset
used to calibrate the imagery, which is discussed in the next section. For the ~1990 time period we used August and
early September 1990 and 1991 Landsat TM imagery. For the ~2000 time period we used August and early
September 1999, 2000 and 2001 Landsat TM and ETM imagery.
Images were registered to the Universal Transverse Mercator (UTM) coordinate system using the NAD 83
datum and nearest neighbor resampling. Careful selection of approximately 30 well-defined, and well-distributed
ground control points (GCPs) resulted in positional accuracy (RMSE) equal or smaller than ± 0.25 pixels, or 7.5 m.
We found the Minnesota Department of Transportation highway map (available in ArcInfo GIS format) provides an
efficient and effective means of easily finding a large number of well-distributed GCPs. Atmospheric correction or
normalization of the Landsat data is not necessary for the regression method described herein. However, it may be
necessary if in-situ data are not available for a particular scene.
Classification Procedures
This section summarizes our image classification procedures; more detail is provided by Olmanson et al.
(2001), and the rationale for the procedures is described by Kloiber et al. (2002ab). ERDAS Imagine, image
processing software, and ArcView, geographical information system (GIS) software, are used for the image
processing steps. One of the most important steps is acquiring a representative image sample from each lake.
Ideally, the sample should represent the center portion of the lake in at least 5 m of water (or twice the SDT
measurement), where reflectance from vegetation, the shoreline, or the lake bottom will not affect the spectral
signature.
The first step is to make a “water-only” image by performing an unsupervised classification in ERDAS
Imagine. Because water features have very different spectral characteristics from terrestrial features, water is put
into one or more distinct classes that can be easily identified. Terrestrial features then are masked creating a wateronly
image. A second unsupervised classification is performed on the water-only image. Average brightness values
from the unsupervised classification of this image are graphed to show spectral signatures of each class. These
signatures along with the location where the pixels occur are used to differentiate classes containing clear water,
turbid water, and shallow water (where sediment and/or macrophytes affect spectral response). Based on this
information classes are re-colored so that vegetation, bottom and terrestrial effects can be avoided when selecting
lake sample locations or areas of interest (AOI). AOIs are the locations where brightness values from the Landsat
image are obtained to develop relationships with measured SDT. For this assessment a polygon layer was used to
help automate the process when delineating AOIs and is discussed in Olmanson et al. (2001). ERDAS Imagine’s
signature editor is used to extract the spectral data from the image for each AOI. All available SDT data collected
within a prescribed number of days (3 to 7 days depending on availability of data for a given image or path of
images from same day) of the satellite image are used for calibration purposes. A data set that includes 20 or more
ground observations per image, spanning a wide range of ambient conditions is considered acceptable. A multiple
regression is performed using log-tranformed SDT data as the dependent variable and Landsat Thematic Mapper
Band 1(TM1) and the TM1:TM3 ratio as independent variables.
In our previous work, lake water clarity maps were created from the regression model by two methods. The first
method used the ERDAS modeler to apply the model to each water pixel. This method creates a pixel-level lake
map (Figure 1). With this map all water pixels are classified and intra-lake variability can be evaluated. The other
method used the brightness values from the AOI in a spreadsheet program to calculate water clarity for each lake.
The data were then linked to a lake polygon layer in ArcView or another GIS program to create a lake-level waterclarity
map (Figure 2). The latter method was used in this study and has the advantage of providing a water clarity
number for each lake that can be used in other analyses or used in a water clarity database.