905 
nnnH 
■n 
■■■■■■■■ 
itions of dns 
Standard deviations of dn values of three 
cover types showing the usefulness of 
sampling: Guildford area UK 
Hean dn TM-Ch 
4 
5 10 15 20 25 30 
Sample size 
Grass'd Type I- 
Parcel G1 
Grass'd Type 2r 
Parcel L1 
Grass'd Type I-Parcel 
G2 
O’ Grass'd Type 2-Parcel 
L2 
* r Bare soil - Parcel B1 
April 1985 
with which the 
sample is not 
nsor is special 
ich different 
reel has been 
p cover. It is 
/hich should 
be incorporated into the strategy of remote sensing 
detection. Those aspects of the rural environment which 
are predictable should be the starting point of such 
strategies. They are: 
• parcel boundaries are relatively static 
• farmers tend to manage parcels to a uniform 
cover 
• farmers tend to raise a limited number of 
crops in a particular region 
The above circumstances are relevant to detection 
strategies with respect to data acquisition and data 
processing. 
5 CONSEQUENCES FOR SENSOR DESIGN FOR 
RURAL LAND COVER DETECTION 
If spatially sampled data, only one tenth or even one 
hundredth as voluminous as total cover, can be shown 
to be as predictive of land cover as total cover it is time to 
pose the question can sensors be designed to acquire 
such economical sampled data? And alternatively can 
on-board processing be arranged to communicate such 
spatially sampled data at an appropriate level? Data 
reduction procedures have been discussed for a decade 
and a half on the assumption that total cover was 
desirable and that data compression would reorganise 
the spectral information or at least reduce the number of 
spectral dimensions. Principal components analysis is a 
good example of such a procedure where it is assumed 
that a satisfactory approach is to reduce the dimensions 
of the spectral data without compromising the spatial 
record. The main purpose of this paper is to question 
whether most major applications in the field of renewable 
natural resources need total cover. Some of the major 
applications reterred to are crop monitoring and the 
survey of the extent of irrigated land. These are the 
applications which are likely to generate the revenue 
which could support the costs of monitoring and they are 
the applications which governments would regard as 
essential for national and regional agricultural 
management. These are the types of data which 
government agencies already spend heavily to acquire. 
On the basis of the evidence in section 3 above in which 
it has been shown that spatially sampled data is as 
effective as total cover for the detection of rural land 
cover in managed environments. It is appropriate, 
therefore, to ask whether it would be possible to 
construct an instrument which would capture data 
according to an appropriate sampling frame. Ideally the 
pixels should be able to be placed at random within each 
parcel but it would be difficult to construct such a sensor. 
Much easier to construct would be a sensor which 
acquires a systematic spatial sample with a regular 
spatial frequency. The disadvantage of this type of 
sample is that a proportion of the pixels would fall on 
parcel boundaries. For example in a five per cent 
systematic sample of the study area ten per cent of the 
sampled 30 metre pixels fell on parcel boundaries. 
However, it was still possible to detect differences in 
crop cover even when these points were discounted. It is 
suggested that the probability of encountering boundary 
pixels should be reduced by decreasing the pixel size to 
ten metres. 
Here it is important to emphasise that the position of the 
field boundaries is known from the topographic record. 
What is required is the ability to overlay the sampled 
spectral data on the static topographic record to 
determine which pixels should be discounted. It is 
presumed that the identification of boundary mixels 
should be achieved by spatial registration but another 
approach which could be useful would be to discount 
tnose pixels which ainertrorn ine mean spectral 
reflectance for the parcel by more than two standard 
deviations. This last is an approach which has yet to be 
tested. 
An ideal solution to the problem of maximising the useful 
data and avoiding the acquisition of information on 
confusing boundary mixels would be the development of 
a sensor which could point accurately to within one pixel. 
If in addition it was possible to programme the sensor so 
that particular pixels were recorded along specified scan 
lines so that boundaries could be avoided then this 
would be the optimum record with respect to the volume 
of data initially acquired and subsequently transmitted. 
The author has no idea whether such a data filtering 
strategy would be feasible in terms of on-board 
programming but is confident that a procedure such as 
this would be an ideal one for the user in most 
circumstances where the agricultural environment is 
heavily managed. 
6 CONSEQUENCES FOR DATA PROCESSING IN 
RURAL LAND COVER DETECTION 
The discussion in the previous section raises some 
important data processing problems. Spatial registration 
is a notoriously demanding process in terms of