values of 
:d by multiple 
rsis could 
led. Only 
; variance 
reamflow, and 
■ameters were 
then a 
character- 
i extracted 
1 and runoff 
tment the 
e equation 
n turn on the 
e optimal 
ations were 
noff not 
s slices and 
artitionings 
of the catchments to be made e.g. Fig. 1. However, 
when the proportions in each classification were 
calculated, the difference between the methods of 
classification became less obvious (Table 1) i.e. 
as soon as the data were lumped the statistics 
became similar, even though the meanings were differ 
ent . 
Secondly, it was observed that many strong correl 
ations occurred between variables. These appeared as 
negative correlations within groups e.g. a high 
proportion of LEVELA1 corresponding to a low propor 
tion of LEVELA5, and also appeared as positive 
correlations across groups e.g. a high proportion of 
FOREST corresponding to a high proportion of LEVELA1. 
The constants a and b were regressed on this 
reduced set of variables, but not significant 
correlations were found. Although this is a sad 
result, it is useful in that it clarifies some issues 
and points the way to possibly happier conclusions. 
The underlying problem is that of continuity of 
the satellite data, It was necessary for the integrity 
of the regression that each set of parameter values 
e.g. CLASSA, LEVELA should be statistically homog 
eneous. The simplest way of ensuring this is to 
evaluate their values from only one image, in which 
case the characteristics of radiation, the calibrat 
ion of sensors etc. will be consistent between each 
catchment. This requires that the catchments studied 
should lie within an area of 185km x 185 km for 
Landsat MSS, or a smaller area for Landsat TM or 
SPOT. This is such a small geographic region that 
there is inevitably a high degree of hydrological 
and environmental similarity between the catchments. 
Thus, the dependent variable does not have sufficient 
variance to make a good regression possible. 
To achieve a higher variance we should need to 
include in the regression catchments which are much 
more dissimilar in their hydrology. To achieve dis 
similarity we should include catchments which are 
remote from one another, but this would mean inter 
preting more than one image, and different images 
will inevitably be of different dates and seasons, 
and different thematic properties. Thus, it appears 
that we should reconsider our spectral characterist 
ics, and use only those which are absolute and 
consistent in range and scale from place to place 
and time to time. The logical conclusion of this 
argument is that we cannot achieve what we set out 
to: either we have insufficient variance for a 
regression model, or we have non-homogeneous data. 
Apart from the problems with regression, we have 
clearly lost much useful information by lumping the 
catchment characteristics. In an attempt to separate 
the two main benefits which we saw in satellite data 
we have lost both. 
Although the regression on constants a and b was 
unsuccessful, useful regression equations were found 
for average yield from the catchments and for mean 
annual flood. Of course, these are very crude 
summary statistics, but the fact that useful regres 
sions were found using 'satellite' parameters in the 
place of conventional characteristics is encouraging, 
and supports our view that satellite data can be use* 
ful if the right model is found. 
5 CONCLUSIONS 
The hydrological characteristics of an area were 
seen to depend on the same features as the spectral 
characteristics. Thus an attempt was made to construct 
an hydrologic model which employed spectral charac 
teristics as its parameters. The attempt failed 
because the variance of the hydrological character 
istics from the area contained in one Landsat frame 
was too small. To achieve greater variance data 
would have to be drawn from more than one image. The 
problems of achieving homogeneity in data from 
different images and different dates appears to be 
insuperable. 
Table 1. Breakdown of classifications - proportion 
of total area in each class 
Classification 
a 
b 
C 
d 
e 
f 
Class 1 
0.62 
0.38 
0.44 
0.57 
0.45 
0.35 
2 
0.22 
0.20 
0.23 
0.16 
0.36 
0.29 
3 
0.15 
0.16 
0.17 
0.14 
0.11 
0.23 
4 
0.007 
0.15 
0.16 
0.13 
0.08 
0.10 
5 
0.12 
- 
- 
- 
0.03 
If satellite data is to prove useful then different 
models will be required. Crude hydrological statis 
tics can be estimated satisfactorily, using regres 
sion models, but to achieve greater sophistication 
it appears that we shall have to turn to the benefits 
of distributed models. 
REFERENCES 
Chidley, T.R.E. and Drayton, R.S. (1985) Visual 
interpretation of standard satellite images for 
the design of water resources schemes. Proc.Int. 
Workshop on Hydrologic Applications of Space 
Technology. International Association of Hydrologic 
Sciences. Florida 1985. In press. 
Drayton, R.S. and Chidley, T.R.E. (1985) Hydrologic 
modelling using satellite imagery. Proc.Int.Conf. 
on Advanced Technology for Monitoring and Process 
ing Global Environmental Data, Remote Sensing 
Society, London. 219-225. 
Gumell, A.M. , Gregory, K.J., Hollis, S. and Hill, 
C.T. (1985) Detrended correspondence analysis of 
heathland vegetations: the identification of 
runoff contributing areas. Earth surface Processes 
and Landforms, 10(4), 343-351. 
Rango, A., Feldman, A., George, T.S. and Ragan, R.M. 
(1983) Effective use of Landsat data in hydro- 
logical models. Water Res. Bull., 19(2), 165-174.