3. Continued halving the time series in this way until
a time span of a week or so was reached. Images
were then viewed sequentially by day, using the
definite ice-off image for reference.
Certain weather events helped and hindered
lake ice classification. Recent snowfall helped
greatly, showing a sharp distinction between ice-
covered lakes and surrounding land (covered with
highly-reflective fresh snow) and open water (having
melted the newly-fallen snow). Cloud shadow
hindered classification, by making lakes appear
much darker than on the preceding cloudless day.
Due to the uncertainty, we classified these
cloud-influenced days as cloudy days.
Some of the lakes were easily identifiable as
snow-covered lakes, but difficult to identify as open
water. They never appeared as dark as others, even
months after ice-off on surrounding lakes. This
phenomenon of unachieved "darkness" varied over
the years for these lakes, and wasn't absolutely
related to lake size, although it was more common
on smaller lakes. A possible explanation is sensor
falloff of the GOES with increasing latitude.
2.5.2 Locational Control
Consistent location of each of the eighty-one
lakes selected for the study was critical. While
interpreting the first few years of images, atlases
containing sufficiently large-scale maps were relied
on to ensure that the correct lake was being
interpreted. After viewing a thousand images or so,
locating lakes became easier, especially in relation to
easily-identifiable lakes and lake-patterns in a
region. Still, we referred to atlases in instances of
any uncertainty.
One potential difficulty in the interpretation
was locating lakes when they had the same
reflectance as surrounding land, and were thus hard
tolocate in the image. This can be seen in Figure 2
(previous page), where Trout Lake is unidentifiable
in some of the images. This was overcome by
overlaying the ambiguous image with a reference
image where the lake is easily identified (ice-on or
ice-off), using the multiple-window viewing capacity
of Imagine. The two (or more) images were spatially
matched using points identifiable on all images,
usually angular shorelines of nearby lakes.
90
This simple overlay of images was possible
due to the geostationary positioning of the GOES
satellite. Rotation or scale changes in the GOES
image resulted in differences of only a few pixels
across the whole image from day to day, making
2-dimensional translation the only necessary image
transformation. However, cumulative changes over
weeks necessitated a local fit (100-400,000 km? at a
time) for overlaying temporally-disparate images.
2.6 Classification Bias
As we progressed with the interpretation,
comparison of interpreted ice-off dates to reference
ice-off dates from ground observation confirmed a
tendency to interpret lakes as ice-off earlier than
they actually were. This is probably due to the
difficulty mentioned previously, in differentiating
between bare ice and open water. This echoes
similar findings by Maslanik and Barry (1987) and
Wynne and Lillesand (1993). Figure 3 and Table 1
(next page) show a comparison for the five lakes for
which ground reference data were available. The
mean absolute difference between interpreted ice-off
and reference ice-off was 3.2 days and the mean
difference was -0.4 days, meaning interpreted ice-off
preceded actual ice-off on average.
N o N a
Mean Difference (Days)
À
-6
1980 1982 1984 1986 1988 1990 1992 1994
Year
Figure 3 Mean Difference between GOES-
Derived and Reference Ice-Off
Dates for Subset of Study Lakes
In light of this bias, we were faced with the
decision whether to correct our interpretation bias
during the interpretation, or “call them as we see
them” and deal with any bias in the statistical
analysis. We chose to do the latter, maintaining our
1992 1993 1994 Abs. Mean Mean
1987 1988 1989 1990 1991
1986
1984 1985
1981 1982 1983
1980
Lake