Geoscience Reference
In-Depth Information
avoids the use of empirical burning effi ciency and vegetation density factors used
in traditional approaches.
Land cover and land-use change
Remote sensing can provide operational information on land-use and land-cover
state and change. Land use describes the human land management practices, while
land cover means the biophysical properties of the land surface. The mean surface
refl ectance at each wavelength recorded by the sensor is called spectral signature.
Many land cover types have characteristic spectral signatures that can be used to
discriminate between them. The more similar two spectral signatures are, the higher
the risk of confusion between land cover classes. There is a general trade-off between
the number of land cover classes (thematic detail) and classifi cation accuracy. The
desire for more classes generally leads to a lower spectral separability between some
of the classes. One way to tackle this problem is to defi ne a hierarchy of land cover
classes, in which the fi rst level of the classifi cation discriminates a low number of
very accurately classifi ed land cover types, while the next level subdivides the classes
further with less accuracy. We can distinguish supervised and unsupervised classifi -
cation algorithms. On the one hand, in a supervised approach, the spectral signa-
tures of each land cover type are defi ned
a priori
either from a digital library of
spectral signatures or by visually identifying areas of known land cover and estimat-
ing the signatures from the image data. In an unsupervised approach such as
k
-
means or Isodata, no prior knowledge of the land cover types in the image is
assumed, and an algorithm automatically cagetorises the image data into a given
number of classes. Unsupervised approaches tend to fi nd better class separability
because they merge and split classes to maximise the difference between their spec-
tral signatures. On the other hand, an unsupervised classifi cation needs careful
interpretation. Since the classes are purely based on spectral characteristics, they do
not necessarily map onto the thematic requirements of a land cover map.
In the UK, land cover mapping from satellite is used in routine environmental
monitoring. Fuller et al. (1994) mapped the entire land area of Great Britain using
Landsat imagery (30-m resolution) as part of the Countryside Survey 1990. The
survey was later repeated in 2000, leading to a new updated map that used an
advanced algorithm (Fuller et al., 2002). Whereas the 1990 map is a pixel-based
map, the UK Land Cover Map 2000 is a polygon database, in which each polygon
or land parcel is linked to a comprehensive attribute data table, containing infor-
mation on the most likely, second most likely land cover class and so on. The maps
have been used by over 500 registered data users and are a prime example of the
operational use of satellite remote sensing. At European level, a standardised land
cover map is also available. The CORINE 1990 and CORINE 2000 land cover
maps with their 44 classes were generated based on amalgamations of nationally
produced land cover maps (e.g., the UK Land Cover Map), but the minimum
mapping units were increased to 25 ha, which limits their use for site-specifi c
habitat management.
It can be hard to reconcile land maps using different remote sensing classifi cation
schemes, since map producers often have their own locally specifi c understandings
what each land cover class actually means in the real world. This subjectivity in
remote sensing image classifi cation and interpretation can lead to substantial seman-
