ages was possible 
ng of the GOES 
es in the GOES 
only a few pixels 
/ to day, making 
necessary image 
tive changes over 
400,000 km? at a 
sparate images. 
he interpretation, 
lates to reference 
ition confirmed a 
>-off earlier than 
bably due to the 
in differentiating 
ter. This echoes 
Barry (1987) and 
ire 3 and Table 1 
the five lakes for 
re available. The 
nterpreted ice-off 
/s and the mean 
nterpreted ice-off 
b 
’ eo 
  
199019921994 
between GOES- 
eference Ice-Off 
Xf Study Lakes 
re faced with the 
nterpretation bias 
| them as we see 
in the statistical 
, maintaining our 
  
1992 1993 1994 Abs. Mean Mean 
1987 1988 1989 1990 1991 
1986 
1980 1981 1982 1983 1984 1985 
-1 
Lake 
  
-1.3 
-1.2 
1.0 
-1.4 
0.7 
27 
-2 
Mendota 
2.9 
-6 
-3 
Monona 
2.7 
Trout 
3.7 
4.2 
3.2 
-1 
Big Trout 
Island 
-3 
5.0 
-5.0 
13 
-1.3 
3.4 
0.2 
2.8 
25 
24 
-0.8 
3.2 
-1.2 
5.8 
-5.4 
40. 30 25 
2.6 
2.0 
2.6 
22 
20. 22 
-1.5 
4.8 
-1.3 
Abs. Mean 
Mean 
0.4 
-1.3 
1.5 
-2.5 
1.8 
  
Comparison of GOES-Derived and Reference Ice-Off Dates for Subset of Study Lakes 
Table 1 
oi 
definition of ice-off as stated in the beginning of this 
section (indistinguishable from a definite ice-off 
day). 
Another obstacle to satellite interpretation of 
ice-off date was cloud cover. In many instances, we 
were faced with a situation where a lake appeared 
ice-on, then a series of cloudy days prevented 
interpretation, and the lake appeared ice-off on the 
next cloud-free day. Average cloud cover prior to a 
visible ice-off date was a mean of 3.7 days for all 
lakes over all years. Individual cloud cover periods 
ranged from 0 to 28 days. Faced with the decision 
whether to interpolate an ice-off date within this 
cloudy period, we chose not to. Our bias toward 
calling ice-off too early was partially mitigated by 
the cloud cover "waiting period." 
3. RESULTS AND DISCUSSION 
Overall, we believe that our visual 
interpretation of the GOES images was a reliable 
method of determining lake ice breakup dates. 
Comparison to available ground-derived ice breakup 
dates revealed a mean absolute difference of 3.2 days 
and a mean difference of -0.4 days — relatively low 
errors given the 1-day temporal resolution of our 
image set. 
While images from sensors with a near- 
infrared spectral band (such as AVHRR) would have 
enhanced the lake ice interpretation, especially in 
distinguishing bare ice from open water, image costs 
would have exceeded our budget. According to our 
reference data, the visible-band GOES images still 
gave us an acceptable degree of accuracy. 
Image interpretation proceeded slowly at the 
beginning of this study, because the necessary 
hardware, software, classification protocol and lake 
set were being designed and/or selected. Time and 
equipment requirements were relatively few once 
these requirements and protocol were established. 
With the use of a high-speed workstation computer, 
the 122 images for each year were processed and the 
81 lakes interpreted in 15-20 working hours. 
The methodological issues we encountered 
would be relevant to other studies of the same 
nature. Basic issues included having enough disk 
space for image processing and storage, and setting 
image boundaries to include selected lakes. More 
complicated issues included selecting easily- 
identifiable and appropriately-sized lakes, dealing 
with sensor falloff and frequent cloud cover in the 
high latitudes, and having a single interpreter for 
consistency.