The UK land cover map 2000: Planning, construction

and calibration of a remotely sensed,

user-oriented map of broad habitats
R.M. Fuller
a,
*
, R. Cox
b
, R.T. Clarke
c
, P. Rothery
b
, R.A. Hill
b
, G.M. Smith
b
,
A.G. Thomson
b
, N.J. Brown
b
, D.C. Howard
d
, A.P. Stott
e
a
Unit for Landscape Modelling, Sir William Hardy Building, Tennis Court Road, Cambridge CB2 1QB, UK
b
Centre for Ecology and Hydrology, Monks Wood, Abbots Ripton, Huntingdon, Cambridgeshire PE28 2LS, UK
c
Centre for Ecology and Hydrology, Winfrith Technology Centre, Dorchester, Dorset DT2 8ZD, UK
d
Centre for Ecology and Hydrology, Lancaster Environment Centre, Library Avenue, Bailrigg, Lancaster LA1 4AP, UK
e
Land Use and Rural Affairs Science Unit, Department for the Environment Food and Rural Affairs,
1/09 Temple Quay house, Temple Quay, Bristol BS1 6EB, UK
Received 1 April 2004; accepted 18 April 2005
Abstract
Land cover map 2000 (LCM2000) is a comprehensive survey of UK broad habitats giving vector digital maps from segment-
based classication of remotely sensed satellite data. This paper examines the inuence of users in designing LCM2000 and the
difculties in applying a user-dened classication. It assesses problems and successes through comparisons with a sample-
based eld survey. These suggest that LCM2000 accuracy at broad habitat level may be around 8085%; however, it was not
possible fully to discriminate errors in LCM2000 from those of the eld survey or from mismatches in scales, resolutions and
survey dates. Calibration generated broad habitat cover statistics fromLCM2000 data to eld survey equivalence. These take full
account of the heterogeneity of a study area, helping to generate accurate statistics, including those at local level where the eld
survey cannot operate effectively. The paper concludes that the comprehensive and extensive coverage from remote sensing
comes closer than alternative methods to meeting users needs. However, it recognises that producers of remotely sensed
information need to understand better the needs of users, and users need to appreciate what the technology can and cannot
deliver. This paper adds some benets of hindsight to the process of communication.
# 2005 Elsevier B.V. All rights reserved.
Keywords: Remote sensing; Land cover; Map; User; Classication; Calibration; Integration
1. Introduction
Land cover map 2000 (LCM2000) is a compre-
hensive survey of UKland cover, part-funded by users.
It updates and substantially upgrades (Smith and
www.elsevier.com/locate/jag
International Journal of Applied Earth Observation
and Geoinformation 7 (2005) 202216
* Corresponding author. Tel.: +44 1223 764379
E-mail address: [email protected] (R.M. Fuller).
0303-2434/$ see front matter # 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.jag.2005.04.002
Fuller, 2000) the land cover map of Great Britain
(LCMGB) (Fuller et al., 1994a), offering improve-
ments in structure, thematic detail (Fuller et al., 2002)
and associated metadata (Smith and Fuller, 2002).
Thematic mapping from remote sensing inevitably
balances the needs of users and the ability of image
data and analysts to deliver. Emissions and reec-
tances of radiation from the earths surface cannot
discern all thematic classes. Users may demand ner
spatial resolution, even mixing resolutions, measuring
ne scale linear features and coarser areal ones.
Classication schemes may confuse land cover and
use, the latter being difcult to discern through remote
sensing. Maps can be made to emulate user needs or
producers might focus on spectrally distinct classes.
The former strategy compromises accuracy, the latter
risks constructing maps which are unt for purpose.
2. Aims
This paper explains the problems faced and
solutions derived in mapping Great Britain. It offers
an account where the both the users and the
producers perspectives are prominent. It outlines
the evolution of methods, the many and varied
inuences which came to bear, and the resultant
products. It describes the process of calibration to
ground reference data collected to user-dened
standards. It gives an insight to inform producers of
similar products elsewhere.
3. Background
Land cover mapping from remotely sensed satellite
images is becoming established (Cihlar, 2000). Much
of the work involves coarse resolution approaches
(Defries and Belward, 2000). Finer scale surveys
include: the US National Land-Cover Database
(Homer et al., 2004); the European CORINE Land
Cover programme (European Environment Agency,
1999); the LGN land cover database for the Nether-
lands (Thunnissen and De Wit, 2000) and LCMGB
and LCM2000. Fuller et al. (1994a) describe the
production of LCMGB. Inter-comparisons of eld and
LCMGB data were made by Wyatt et al. (1994),
Cherrill et al. (1994, 1995) and Fuller et al. (1998).
Several papers give a background (Comber et al.,
2003), context (Fuller et al., 2002) and description for
LCM2000 (Fuller et al., 2002, 2004; Smith et al.,
2000). While accuracy assessment is widely applied,
methods vary considerably. Reviews by Congalton
(1991) and Foody (2002) give more general context
for the present work. Example applications over large
area studies are given by Edwards et al. (1998), De Wit
and Clevers (2004) and Van Oort et al. (2004).
3.1. LCMGB
LCMGB is a classication of spectral data from
earth resources satellites (Fuller et al., 1994b). Landsat
Thematic Mapper (TM) data, mostly of 19881989,
were registered to the British National Grid (BNG)
with 25 m output pixels. Red, near-infrared (NIR) and
middle infrared (MIR) data from summer and winter
were combined into six-band composite images
(Fuller et al., 1994b). A maximum likelihood
classication (Schowengerdt, 1997) labelled each
pixel in each scene with one of 25 classes (Fuller
et al., 1994a). Amosaic of classied scenes covered all
of Britain.
The Institute of Terrestrial Ecology (ITE), later the
Centre for Ecology and Hydrology (CEH) produced
LCMGB with science budget funds and a grant from
the British National Space Centre (BNSC). Concur-
rently, the then Department of the Environment (DoE)
was planning to part-fund ITE in the sample-based
Countryside Survey (CS) 1990 (Barr et al., 1993). The
comprehensive if generalised coverage of LCMGB
would clearly complement the sample-detail of the
CS1990 eld survey. DoE gave funds to integrate
LCMGB into CS1990, allowing inter-comparison
with the eld survey (Fuller et al., 1998). Since
production, LCMGB data have been licensed to over
500 users, researchers, policy makers and commercial
organisations, with wide ranging uses.
LCM2000 had the benets of hindsight that
followed production and use of LCMGB, but there
were still lessons to learn. Given the scope for
comparisons to identify changes or the alternative
potential to upgrade the product, there were newissues
to consider. Users variously demanded greater
thematic accuracy, increased thematic detail or
improved spatial resolution. LCMGB had been
completed as geographical information systems
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 203
(GIS) started to become widely used. Users were
hungry for data of many sorts. Many were inexper-
ienced in issues of data formats, quality, their use and
tness for purpose. Thus, users needed advice on what
could and could not be done with the LCMGB data.
However, there were misunderstandings and misuses.
Many users failed to accommodate the 1520% errors
in a map which is 8085% accurate (Fuller et al.,
1998). Sometimes, the informed buyer of the data was
not the user: legitimate and unlicensed copies were
made, the data misused, and users publicly criticised
aspects of data quality, despite the warnings.
Integration of eld and LCMGB data was limited.
The CS1990 launch publication (Barr et al., 1993)
presented independent national cover statistics, which
conicted due to thematic and spatial differences. In
comparing results, Fuller et al. (1998) noted the
products complementarity but advocated better
integration to exploit that. They proposed a much
closer integration of LCM2000 with its concurrent
eld survey. This paper describes two stages of that
process: rst, adoption of a common classication
scheme; second, development of methods to inter-
calibrate outputs.
3.2. Planning LCM2000
LCM2000 was to be more widely useable than
LCMGB, with improved accuracy and extended
thematic detail. A common classication, a coincident
timetable and a similar data structure were adopted to
improve integration with the eld survey. There was
little prospect, logistically or economically, of
increasing spatial resolution. However, there was
scope to improve spatial characteristics and data
formats; and to enhance substantially the metadata
retained in nal outputs.
LCM2000 needed user funding. The Secretary of
State for the Environment in principle committed the
government to a decennial repeat of the CS at the
CS1990 launch in 1993. The new government of 1997
brought this forward to 1998. In planning CS2000, the
DoE and its successor, the Department of the
Environment, Transport and the Regions (DETR),
drew together a consortium of users to fund CS2000.
Conservation agencies (the Countryside Council for
Wales, and Scottish Natural Heritage) and the
Environment Agency (EA) joined; the Scottish Ofce
and Welsh Ofce, later the devolved administrations
of Scotland and Wales, became members; also the,
then, Ministry of Agriculture, Fisheries and Foods
(MAFF). With an ITE contribution, there were
enough funds to map Britain. Later, the Northern
Ireland Environment and Heritage Service and the,
then, Department of Agriculture for Northern Ire-
land became partners, extending coverage to the
entire UK. ITE then had to meet the demands of 10
partners.
3.3. Methodological improvementssegment-
based mapping
Prior to LCM2000, research was undertaken to
improve outputs. CEH, the Ordnance Survey and the
Department of Geography in the University of
Cambridge collaborated to develop methods for
vectorraster mapping (Smith et al., 1997). The
BNSC, backed by DETR, part-funded the work.
Procedures were developed for segmenting satellite
images to give vector outlines (Devereux et al., 2004)
and segments were classied by their spectral
characteristics. LCM2000 adopted these procedures.
The new structure would preclude direct comparisons
with the raster LCMGB(Comber et al., 2003, 2004a,b)
compromising change detection. However, a rene-
ment of the cover classes already hindered this
process. As unrealistically high levels of precision are
needed for reliable change mapping (Fuller et al.,
2003) a product upgrade was deemed the priority.
3.4. The broad habitat classication
A broad habitat classication was developed for
reporting under the UK Biodiversity Action Plan (UK
Biodiversity Steering Group, 1995). More narrowly
dened priority habitats have specic conservation
targets under the European Habitats Directive
(Council Directive 92/43/EEC). Monitoring of broad
and priority Habitats contributes to the UN Conven-
tion on Biological Diversity (UNEP, 1992). DETR,
responsible for monitoring, wanted LCM2000 to
contribute. In Scotland, Wales and Northern Ireland
the devolved administrations inherited these respon-
sibilities. The conservation agencies, charged with the
process of monitoring, wanted use of the scheme. It
met the EAs more general needs. MAFF accepted the
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 204
classication with supplementary information on
arable crops.
Jackson (2000) determined that the broad habitats
would be: comprehensive, exclusive (with a once-
only t), structured (a framework for priority
habitats), nested (priority habitats tting into only
one broad habitat), measurable (a distinct surface area
with clearly characterised features) and consistent (not
sub-dividing some ecological units more nely than
others). These are desirable features of any classica-
tion. Accordingly, LCM2000 adopted widespread
examples of terrestrial, freshwater and coastal broad
habitats (Table 1). However, the classes did they
always meet the prescribed criteria, and they were not
designed for remotely sensed mapping. Some were
dened by plant species complement, others by
substratum, neither evident from satellite images.
Some habitats are below the spatial resolution of the
imagery. Others are too rare or localised to provide
adequate reference sites for extrapolation. Some
denitions include elements of land use, not dis-
cernible from images. Other users also wanted classes
beyond the broad habitats: in particular, the con-
tinuous urban and discontinuous suburban classes
have LCMGB and arable crop types.
4. Image analyses
4.1. Images
LCM2000 used combined summer and winter
images from enhanced TM, TM, and the linear self-
scanning sensor (LISS) (Fuller et al., 2002). The target
summer period was the main growing season for
arable crops; to match the eld survey, this was to be
the summer 1998. The target winter period actually
ran from autumn to spring; to relate to the 1998
cropping season, these images would come from
winter 19971998. Compromises on the seasonal
period and the choice of years were allowed where
target dates could not be met.
Under ideal conditions, UK cover would have
taken 46 satellite scenes. In the event, the years around
1998 were bad for imagery and LCM2000 used 79
scenes. Wintersummer composites gave 78 combi-
nations (compared with 32 for Britain in LCMGB
(Fuller et al., 1994b)). Eighty-four percent of the UK
was mapped from wintersummer images with 9%
from summer-only and 6% winter-only cover. Only
23% of the UK was mapped from images recorded in
the target seasons of 19971998. No other 19971998
wintersummer composites were useable. However, a
further 22% of the UK used one 1998 dataset
combined with images from other years. Hence,
45% of the UK was mapped with a contribution from
1998 data. The combination of summer data with
winter cover from another cropping year helped most
class distinctions, but caused confusion in its
composite view of two crops. In all, 54% of the UK
used data entirely from non-target years.
The need to classify 61% extra images reduced the
time spent per scene. Images from mixed sensors and
adjacent sensor paths fragmented the map and reduced
productivity. LISS data, with a coarse (72 m) MIR
band and band striping due to sensor decay,
contributed to 22% of the map, reducing quality
and demanding extra effort. Segmentation of compo-
site images produced no segments on single-date
sections, which were treated separately. As the
classication process was iterative, the extra workload
reduced the number of iterations and affected quality.
Contextual corrections needed tailoring to each scene-
combination. Thus, the 78 scene-combinations
demanded highly individualised treatments.
The worst impacts came from using images
outside the target seasons. In 19972001, the best
summer weather was May, before deciduous trees
were fully in leaf and spring-sown crops had
germinated. June and July provided little imagery.
Late summer images post-dated much of the arable
harvest, precluding crop distinctions and causing
confusion between arable elds and bare ground. On
winter imagery, a patchwork of cloud and shadow
often obscured large areas. In mid-winter, low light
levels sometimes only used 510% of a sensors
intensity range, giving little discrimination between
objects and no data at all in shaded facets of terrain.
In 16% of the UK, lack of winter imagery
exacerbated the impacts of off-season summer cover.
Image pre-processing (Fuller et al., 2002) aimed to
reduce the problems: sharpening of LISS MIR data
using visible and NIR bands, atmospheric haze
correction, cloud and shadow removal; snow mask-
ing in winter scenes; and correction of terrain-
induced differential illumination.
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 205
R
.
M
.
F
u
l
l
e
r
e
t
a
l
.
/
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
A
p
p
l
i
e
d
E
a
r
t
h
O
b
s
e
r
v
a
t
i
o
n
a
n
d
G
e
o
i
n
f
o
r
m
a
t
i
o
n
7
(
2
0
0
5
)
2
0
2

2
1
6
2
0
6
Table 1
Broad habitats and their relation to LCM2000 target classes, subclasses and class variants; also a simplication of all these into aggregate classes; hard lines between classes shows
mostly reliable distinctions; dotted lines are shown where distinctions less certain
N.B. the nomenclature may be abbreviated to keywords in the text and other tables.
4.2. Segmentation
Segmentation identied uniform areas in images
as the basis for LCM2000 polygons (Smith and Fuller,
2000; Smith et al., 2000). Edge-detection identied
boundaries; region growing used seed points selected
to avoid boundary pixels (Devereux et al., 2004). A
threshold was set for a eld-by-eld segmentation
whilst also separating urban and suburban areas and
subdividing heterogeneous semi-natural zones. Intel-
ligent generalisation methods excluded segments
8 pixels (0.5 ha). While this inevitably discarded
real information, it reduced noise and gave a product
more like conventional mapping. Finally, raster-to-
vector conversion created vector versions in the GIS
database. These recorded features such as lakes, elds,
woods and towns as GIS objects.
4.3. Broad habitats and LCM2000 classication
LCM2000 is a classication of spectral data.
Spectral classes (Kershaw and Fuller, 1992) can be
combined into thematic components, which can be
aggregated into various classications (Table 1).
While LCM2000 aimed to map broad habitats, in
practice the target classes were the nearest matches for
consistent and accurate mapping. Subclasses were
added to give the full complement of broad habitats,
using external contextual data to help distinction.
They were mapped consistently throughout the UK,
but with lower accuracy than target classes. The
building blocks of all classes were the 72 class
variants, the thematic aggregations of the spectral
classes, recognised where possible but not necessarily
consistently (e.g. crops once harvested could not be
distinguished).
Classication (Fuller et al., 2002) used a repre-
sentative sample of training segments as ground
reference to determine the spectral characteristics of
key cover types. A maximum likelihood classication
(Schowengerdt, 1997) compared each segments mean
reectance statistics with the training data. Only the
core (non-edge) pixels of segments and training
segments were used. The procedure recorded the most
likely class and its probability together with such
details for the second- to fth-choice classes, usually
covering >90% of the probability range. Contextual
procedures (Fuller et al., 2004) rened distinctions.
Scene-specic corrections, made immediately after
classication, identied segments classied with low
condence or classes out of context (e.g. coastal
habitats found inland) and re-labelled them appro-
priately. Then, individual classied scenes were
mosaicked and cloud-holes patched to complete
coverage in 100 km 100 km sections. Further
contextual coding, using external data, extended the
class distinctions for some habitats.
The vector dataset comprises 6.6 million segments,
held as 100 km 100 km sections in ArcView
Shapeles. A key characteristic of LCM2000 is the
retention of detailed metadata (Smith and Fuller,
2000). In the original GIS, each segment carries
attribute data that describe: shape, size and location;
source images and dates; broad habitat, target class,
subclass, class variant and spectral class; probabilities
for the top ve spectral classes; details of contextual
corrections; pixel-based scores of within-segment
heterogeneity; and LCMGB and CORINE 1990
classes.
5. Calibration of LCM2000 to eld survey
broad habitats
LCM2000s comprehensive if generalised cover-
age were calibrated against the eld surveys detail to
give outputs that neither survey could supply alone.
5.1. Field survey data
The eld survey recorded a stratied random
sample of 1 km
2
(Haines-Young et al., 2000). There
were 40 strata and 569 squares in Great Britain, 519
recorded in 1998, the rest in 1999, Digital data were
not available for Northern Ireland, precluding use in
calibration. Surveyors recorded details including land
cover; and use; plant species lists; and a range of data
on boundaries and linear features (Haines-Young
et al., 2000). A primary coding of cover, with
qualifying secondary codes, was applied to parcels
on OS 1:10,000 scale maps. Additional boundaries
were added as necessary (e.g. in semi-natural
mosaics). The minimum mapable unit was 0.04 ha.
Objective evaluations of primary/secondary codes and
plant species data generated broad habitat labels for
each parcel. Haines-Young et al. (2000) give habitat
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 207
cover estimates extrapolated from the sample, using
the generalised stratication. Though national esti-
mates are good, local estimates are generally
unrealistic.
The eld survey had been established in 1978. The
format was maintained in successive surveys to record
change. There was little scope to alter, extend or re-
phase the survey for integration with LCM2000. Thus,
LCM2000 adopted a common timetable, the same
broad habitat classication and a shared vector data
structure to improve its compatibility with the eld
survey.
The eld data are not ground truth: independent
checks (Prosser and Wallace, 1999) showed 88%
repeatability for the primary codes which essentially
determined eld broad habitat labels. Habitat bound-
aries in unenclosed land proved impossible to map
repeatably. The eld survey adopted the 1990 outlines
and codes (with equal deciencies), unless there had
been obvious changes. Even the check survey used the
1990 boundaries, so their quality was not assessed.
Thus, the 88% correspondence over-states eld survey
accuracy.
5.2. GIS comparisons
Arc/Info les, labelled with broad habitats, were
generated for all 569 eld survey squares and
equivalent LCM2000 map-sections. They were used
in:
Per-pixel comparisons between eld survey and
LCM2000 maps: a direct overlay, accepting the
structural differences between the datasets;
Per-segment comparisons, where labels in
LCM2000 segments were compared with each
segments dominant class according to the eld
survey;
Per-parcel comparisons, where labels in eld
survey land parcels were compared with the
dominant class according to LCM2000.
Correspondences were calculated at broad habitat
level. For simplicity, the analyses (per-pixel, per-
segment and per-parcel) used a raster-GIS approach to
score correspondence. Field survey parcels and
LCM2000 segments were sampled onto a 2.5 m grid,
ne enough to record the continuously variable
outlines of the eld parcels with correspondences
matching vector-based results. Scoring recorded the
160,000 sample-pixels in each 1 km
2
. The results
for parcels and for segments were aggregated and
weighted by their areas.
5.3. Inter-calibration procedures
The basic data for the calibration were the
individual calibration matrices for each 1 km
2
. For
any one square k, element M
k
i j
in row i column j of the
calibration matrix M
k
denotes the number of pixels in
the square classied as LCM2000 cover type j which
were classied as eld survey broad habitat cover
type i.
Calibration was initially done separately for each
stratum. First, matrix M was calculated where:
M M
1
M
2
. . . M
s
(1)
is the sum of the calibration matrices for the s sample
eld survey squares in the stratum.
The procedure then calculated the average calibra-
tion matrix P for the stratum, whereby element P
ij
of
matrix P is dened by:
P
i j

M
i j
P
i
M
i j
(2)
and equal to the proportion of the total number of
pixels in column j of matrix M which are in row i.
For the region of interest, let L denote the vector of
known areas for each of the LCM2000 cover types in
the stratum. Then, in that region of interest, the vector
F gives estimates of the area of each of the eld survey
cover types in the stratum by:
F P L (3)
The estimated total area (A
j
) of eld survey broad
habitat type j in the region of interest was obtained by
summing its calibration-estimated area, F
jk
, in each
stratum k over all the m strata:
A
j
F
j1
F
j2
. . . F
jm
(4)
To calculate condence limits for the calibrated
estimates, bootstrap samples (Efron and Tibshirani,
1993) were obtained by random simulated resampling
of the matrices M
1
, M
2
,. . ., M
s
, with replacement,
separately for each stratum. For each bootstrap
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 208
sample, the average calibration matrix P was
calculated for a stratum using Eqs. (1) and (2), and
then Eqs. (3) and (4) were used to derive the bootstrap
sample estimate of the area of each eld survey broad
habitat. The 2.5 and 97.5 percentile values of the
bootstrap estimates provided 95% condence limits.
The approach was extended to allow for stratication.
The eld survey squares were selected according to a
stratied random sampling scheme based on the 40
strata. The national estimates of eld survey values
were calculated by areally weighting the estimates
derived separately for each stratum:
FS
A
1
F
1
A
2
F
2
. . . A
n
F
n
A
1
A
2
. . . A
n
(5)
where A
i
is the stratum area and F
i
is the vector of the
proportions in each land cover type of the ith stratum.
Each F
i
was re-calculated for each bootstrap sample
by randomly sampling from the set of confusion
matrices for the given stratum (Eq. (1)). Trials showed
that 1000 bootstrap samples were sufcient to calcu-
late the 95% condence intervals. Bias-corrected per-
centile limits (Efron and Tibshirani, 1993) removed
any bias that arose because the true parameter value
was not the median of the distribution of estimates.
5.4. Correspondences
Table 2 gives summary correspondences, with 95%
condence limits for the three comparisons. Per-pixel
correspondence was lowest, recording every spatial
difference between the eld survey and LCM2000,
even where these reect inherent and appropriate
characteristics of the maps (e.g. mismatches in outlines
basedonLCM2000image pixels, alsodifferences inthe
minimum mappable unit (MMU)). Estimated per-pixel
correspondence in Britain at broad habitat level is
54% (95% range: 5356%), in England and Wales the
match is 60% (range: 5862%), in Scotland it is just
44% (4047%).
LCM2000 segment labels, compared with those
from eld survey, give higher correspondences: 58%
(5760%) in Britain, 64% (6266%) in England and
Wales and 47% (4350%) in Scotland. The measure
shows how the spectral-classication of segments
fared, compared with what a eld survey of the
segments would have recorded. It accommodated
LCM2000s larger MMU.
Per-parcel correspondence gave the best match:
62% (6064%) correspondence at broad habitat-level
in Britain, 69% (6772%) in England and Wales and
just 48% (4452%) in Scotland. This assessment
accommodated most of the structural differences
between the surveys. It is a measure of how effectively
the LCM2000 class labels could be transferred to
conventional vector maps.
5.5. Assessments at broad habitat level across
Britain
Britains Broadleaved woodland cover is 6.7%
from LCM2000 or 6.6% by eld survey. An overlap of
just 49% relates mainly to boundary differences due to
scale, resolution and LCM2000s underlying pixel-
based structure. Coniferous woodland, mostly planted
and in larger blocks, covers 5.6% on LCM2000 and
6.2% by eld survey, with 82% correspondence; the
eld value includes newly planted and felled areas,
effectively recording a forestry land use class.
Arable and horticultural land covers 24.7% on
LCM2000 or 23.6% by eld estimates with 89%
overlap. Improved grassland covers 21.8% on
LCM2000 or 24.6% by eld survey with 73% overlap.
Apparent inter-confusions relate partly to rotation
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 209
Table 2
The correspondence (percent) between LCM2000 and the eld
survey for sample 1 km
2
in Great Britain, England and Wales
combined and Scotland, measured per-pixel, per-segment and
per-parcel: correspondences are weighted estimates with 95% con-
dence limits, based on 569 1 km
2
in 40 strata, calculated at broad
habitat level excluding linear features
Correspondence
(%)
95% Condence
limits
Lower Upper
Great Britain
Per-pixel 54.1 54.0 54.1
Per-segment 58.2 61.5 61.6
Per-parcel 61.5 58.1 58.2
England and Wales
Per-pixel 59.9 59.8 59.9
Per-segment 64.4 69.3 69.3
Per-parcel 69.3 64.3 64.4
Scotland
Per-pixel 43.6 43.5 43.7
Per-segment 47.0 47.4 47.7
Per-parcel 47.5 46.9 47.2
farmingin squares where survey years differed. Though
misclassications are relatively small, the habitats are
extensive, contributing signicantly to total error.
LCM2000 confusion between Arable and Built up
land is a problem where part-grown, ripened or harves-
ted crops overlap spectrally with vegetated suburbs.
Discrimination of Improved and semi-natural grass-
lands is difcult: improvement is a continuous process;
abandonment can give the spectral character of semi-
natural swards; and reversion is possible (Fuller, 1987).
Field surveyors relied on indicator species. Thirteen
percent of eld survey Improved grassland is recorded
by LCM2000 as semi-natural. No consistent spectral
characteristic separates Neutral, Calcareous and Acid
grasslands. A 1 km grid-based acid-sensitivity map
(Hornung et al., 1995) approximated such distinctions;
results poorly matched the eld survey. Bracken, a
subclass identied for broad habitat mapping, went
undetected on early summer imagery, when bracken
cover was scant.
The Fen, marsh and swamp habitat is . . .
characterised by a variety of vegetation types . . . on
. . . waterlogged . . . soils (Jackson, 2000). The eld
survey identies much more of this (2.5%) than
LCM2000 (0.1%). It included rush-pastures, which
LCM2000 treated as Acid grassland. This distinction
ledthe eld survey to record much more Fen, marshand
swamp in 1998 than in 1990, and raised questions over
the eld interpretation (Haines-Young et al., 2000).
A range of habitats occurs in unenclosed upland
where the eld boundaries default to 1990 outlines. In
such mosaics, even a perfect LCM2000 would not
match the eld survey. There were also denition
problems. Jackson (2000) states that the Bog habitat
. . . resembles wet or dry dwarf shrub heath . . . and
adds that Peat depth, although somewhat arbitrary, is
used as the primary criterion . . . stating that . . . peat
greater than 0.5 m deep is classied as bog . . .. A
geological map showing peat drift >0.5 m was used to
determine context, recoding LCM2000 Bog and heath
as appropriate. This proved a very conservative picture
of peatland: LCM2000 shows 11.3% cover of heath
against eld estimates of 6.7%, and 2.2% cover of Bog
(much less than suggested by Reid and Quarmby
(1997)) against 10.0% cover by eld survey.
Inaccessible Montane habitats were seldom visited
in LCM2000 eld reconnaissance. Vegetated ground
>600 m altitude (Ratcliffe and Thompson, 1988) was
treated as Montane. LCM2000 records 1.7% cover
against the eld estimate of 0.2%. While the eld
survey may have difculties in estimating the cover-
age of such rare habitats, there are questions over
LCM2000s altitude-based distinction. The Inland
rock habitat includes natural (mostly upland) and
articial surfaces. LCM2000 Inland bare ground
included both but may erroneously have incorporated
some bare arable land. Unlike the eld survey, it
appropriately mapped bare ground in the urban
context. Overall, it records four times as much bare
ground (0.9%) as the eld survey (0.2%).
It is clear from the above that LCM2000 upland
classes (Dwarf shrub heath, Bog, Montane habitats
and Inland bare ground) do not exactly match broad
habitat denitions and the eld survey poorly records
upland mosaics. It is not surprising that the maps
differ. An Aggregate Mountain, heath and bog class
(Table 1) gives a much closer 79% correspondence,
and coverages of 16.1% by LCM2000 or 17.2% by
summing eld survey estimates.
Water (inland) on LCM2000 is an aggregation of
the Standing open water and canals and Rivers and
streams habitats, with no attempt to distinguish
standing from owing water. LCM2000 and eld
estimates for Water area are both 0.9%.
Built up areas are mapped differently by eld and
LCM2000 surveys. The former only mapped squares
with <75% urban land, and then treated Built up areas
and gardens as a generalised mosaic. LCM2000
distinguished open spaces >0.5 ha within urban areas.
The eld estimate suggests Built up areas cover in 6.0%
of Britain while LCM2000 records 7.0% cover overall.
Supralittoral rock and sediment, and Littoral rock
and sediment are often below the resolution of
LCM2000. LCM2000 recorded the tidal state at the
time of imaging, while the eld survey used mean low
water on OS maps. The 0.4% total coverage
recognised by LCM2000 differed from eld estimates
of 1.8%, but neither provides nationally consistent
estimates (Haines-Young et al., 2000).
5.6. Correspondence and accuracy?
The correspondence between LCM2000 and the
eld survey is not a measure of LCM2000s accuracy as
the eld survey does not provide ground truth, and
mismatches due to scale, map-structure and differences
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 210
R
.
M
.
F
u
l
l
e
r
e
t
a
l
.
/
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
A
p
p
l
i
e
d
E
a
r
t
h
O
b
s
e
r
v
a
t
i
o
n
a
n
d
G
e
o
i
n
f
o
r
m
a
t
i
o
n
7
(
2
0
0
5
)
2
0
2

2
1
6
2
1
1
Table 3
Broad habitat cover statistics (km
2
and % cover) for Great Britain: eld survey estimates (with 95% condence limits) were extrapolated from 569 sample squares; they are compared
with equivalent data, uncalibrated and calibrated, from Land Cover Map 2000 per-pixel data; LCM2000 estimates outside the eld survey condence limits are italicised
Broad habitats Field survey LCM2000
Cover
estimate
95% Condence
limits
Estimated
% cover
Uncali-
brated
total
% Cover
(uncali-brated)
Calibrated
sample
mean
Bias corrected 95%
condence limits
% Cover
(cali-brated)
Lower Upper Lower Upper
Broadleaved, mixed and yew woodland 14708 12837 16707 6.6 15259 6.7 16125 14679 17019 7.0
Coniferous woodland 13738 10766 16786 6.2 12879 5.6 12995 11677 14079 5.6
Arable and horticulture 52489 48280 56640 23.6 56638 24.7 54743 52433 57300 23.7
Improved grassland 54816 50720 58822 24.6 50024 21.8 59200 56811 61871 25.6
Neutral grassland 6126 5125 7219 2.8 10770 4.7 6674 5850 7525 2.9
Calcareous grassland 647 216 1217 0.3 10647 4.7 533 215 899 0.2
Acid grassland 12955 10950 15166 5.8 14483 6.3 12337 10699 14398 5.3
Bracken 4389 3458 5402 2.0 1892 0.8 3840 3140 4372 1.7
Dwarf shrub heath 14870 12256 17642 6.7 25788 11.3 14048 11909 16026 6.1
Fen, marsh and swamp 5472 4311 6738 2.5 197 0.1 5273 4196 5846 2.3
Bog 22183 19173 25389 10.0 5134 2.2 22578 21037 26048 9.8
Standing open water and canals 1903 972 3103 0.9 2095 0.9 1709 1433 1845 0.7
Montane habitats 491 86 1141 0.2 3971 1.7 606 117 1399 0.3
Inland rock 555 354 791 0.2 2163 0.9 639 347 877 0.3
Built up areas and gardens 13308 11131 15621 6.0 16132 7.0 17652 16597 19812 7.6
Supralittoral rock 1903 972 3103 0.9 19 0.0 782 582 858 0.3
Supralittoral sediment 770 593 969 0.3 166 0.1 341 244 425 0.1
Littoral rock 1 0 4 0.0 55 0.0 0 0 1 0.0
Littoral sediment 1381 609 2303 0.6 637 0.3 977 569 1140 0.4
in survey-date are not errors. LCM2000 segments show
a basic correspondence of 62% with eld data in per-
parcel comparisons at broad habitat level. This rises to
64% allowing for the eld surveys generalisation of
built up areas. As correspondence cannot logically
exceed the 88%repeatability of the eld survey primary
coding, LCM2000 is scoring 72% of potential. The
greatest cause of the difference lies in unenclosed semi-
natural uplands where LCM2000 distinguishes Bog and
heath differently and where eld surveyors were unable
accurately to map vegetation patterns. If we take the
eld survey approach, scoring the agreement in
proportional cover of upland classes, an apparent
success of just 33% in LCM2000s mapping of upland
could be nearer 60%; as upland comprises 30% of the
mapped area, the overall success of LCM2000 might be
10% higher than the quoted 72% of potential. Other
differences also look wrongly like errors in LCM2000:
54% of LCM2000 was mapped using images of
different year to the eld survey. Mostly because of
rotation farming, squares surveyed in the same year
gave 6%closer agreement than average. Had the image
and eld survey dates invariably matched, correspon-
dences would clearly have been higher.
5.7. Calibrated statistics
Calibrated broad habitat statistics for Britain
(Table 3) are based on LCM2000 per-pixel compar-
isons (per-parcel and per-segment results are near
identical). Cover values are close to eld estimates
(Haines-Young et al., 2000) and mostly within the
eld survey condence limits. Out of 11 habitats
where uncalibrated LCM2000 estimates were outside
the eld survey condence limits, 9 came within the
limits after calibration. Notable changes occurred in
Neutral grassland, Calcareous grassland, Bog and
Dwarf shrub heath estimates. Two minor exceptions
were the Supralittoral sediment habitat (for reasons
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 212
Fig. 1. Calibrated LCM2000 broad habitat cover estimates (km
2
)
plotted against equivalent eld surveyed cover estimates for Great
Britain; error bars show bootstrapped 95% condence limits.
Fig. 2. Calibrated LCM2000 broad habitat cover estimates (km
2
) plotted against equivalent eld surveyed cover estimates for England and
Wales combined; error bars show bootstrapped 95% condence limits.
which are unclear) and Built-up areas (where the eld
estimate includes an uncalibrated adjustment, not
updated since 1990, for unsurveyed urban land (http://
www.defra.gov.uk/wildlife-countryside/cs2000/08/
01.htm)). Calibration works as well for separate
country estimates: when calibrated LCM2000 esti-
mates are plotted against eld survey equivalents
(Figs. 13), the slopes are near to unity (1.067, 1.080
and 1.005) and the intercepts close to the origin,
showing there is little bias in the estimation. R
2
values
of 0.995, 0.996 and 0.996 imply a close linear
relationship. It is also evident that condence limits of
calibrated LCM2000 estimates are tighter than those
from the eld survey.
6. Discussion and conclusions
LCM2000 is rst full survey of UK land cover at
such spatial and thematic resolutions. The product is
available under licence to academic, policy-based and
commercial users. It records a hierarchical classica-
tion, broadly at a eld-by-eld scale, offering the only
comprehensive mapping of broad habitats throughout
the UK. The vector-GIS database provides exceptional
levels of detail for a remotely sensed product.
LCM2000 clearly has many features in common with
other large-area studies (e.g. European Environment
Agency, 1999; Thunnissen and De Wit, 2000; Homer
et al., 2004). However, there are characteristics in the
form of outputs and particularly in their calibration
against the detailed eld survey, which give LCM2000
unique qualities.
The eld survey gave unmatched levels of detail for
a nationally representative, stratied random sample
covering the broad habitats and including a diverse
range of point, linear, areal and oristic data (Haines-
Young et al., 2000). It generated extrapolated cover
estimates for all habitats across Britain. However, the
relatively small size of the sample and the coarseness
of the stratication precluded realistic estimations at
local levels and gave wide condence limits for the
cover of rarer habitats, even at national level.
The eld data proved inadequate to validate fully
the comprehensive national survey: it was hard to
better some outputs from remote sensings synoptic
overview. As with conventional mapping generally,
the eld survey essentially presents a caricature of
the real world: complex land cover patterns, con-
tinuously variable in space and time, are articially
recorded as discrete features, separated by sharp,
unbroken lines; parcels are shown as unambiguously
uniform; some features are portrayed at enhanced
scales with selective omissions of others. The picture
from remote sensing may lie closer to the truth.
However, users dislike the complexity of pixel-based
outputs, which can be difcult to interpret and analyse.
It was therefore reasonable to adopt vector mapping in
LCM2000, though it may have incorporated decien-
cies which hitherto have been accepted for pragmatic
reasons of manual cartography rather than for good
reasons linked to applications. In reality, the inten-
sively used landscapes of Britain have a structure not
unlike conventional maps, with discrete boundaries
around elds of essentially uniform crops; even urban
and many semi-natural areas are constrained by
ownership boundaries within which, under a broad
thematic classication, the contents are effectively
uniform. Thus, segment-based mapping proved
suitable for LCM2000. It was also particularly
effective for collecting, storing and analysing the
diverse metadata generated for, and associated with,
segments. Per-pixel measures for each segment
captured some of the heterogeneity masked by the
segmentation.
Comparisons between LCM2000 and the eld data
highlighted LCM2000s problems in spectral mapping
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 213
Fig. 3. Calibrated LCM2000 broad habitat cover estimates (km
2
)
plotted against equivalent eld surveyed cover estimates for Scot-
land; error bars show bootstrapped 95% condence limits.
of some broad habitats. An putative solution, the
contextual analyses, relied on datasets which also
proved inadequate. Even with hindsight, it is hard to
identify better ways of mapping all the habitats.
Unusually poor weather contributed substantially to
the difculties. Given typical weather, hence better
image quality in appropriate seasons and years, the
problems would have been much less. It is unfortunate
that the all weather capability of synthetic aperture
radar (Henderson and Lewis, 1998) was deemed
unsuitable for operational use in broad habitat
mapping.
The complementary qualities of LCM2000 and the
eld survey gave important capabilities: rst, the
potential to indicate approximate LCM2000 map
accuracy; second, an indication of which broad
habitats were most and least reliable on LCM2000;
third, the scope to calibrate LCM2000 cover to eld
survey standards; fourth, the ability to generate habitat
cover statistics from calibrated LCM2000 data; and
fth, the scope to check eld survey extrapolations
using summary cover data from the calibrated
LCM2000.
A direct evaluation of LCM2000 precision is not
possible. Yet, users inevitably demand some indica-
tion of accuracy. Alogical interpretation of observed
differences suggests that LCM2000 could be 8085%
accurate at broad habitat level. However, it was not
possible fully to discriminate errors in LCM2000 from
those of the eld survey or from mismatches in scales,
resolutions and survey dates. At target class level
accuracy may be 5% better. At aggregate class level,
correspondences are about 12% higher, and most
LCM2000 results are reliable, within constraints of
scale and resolution. These results differ little from
those of LCMGB (Fuller et al., 1998). However,
LCM2000s strengths lie in its upgraded vector
structure and its detailed database.
Calibration gave broad habitat cover statistics from
LCM2000 to eld survey equivalence. At national
level, results closely match those extrapolated from
eld samples, implying that the samples adequately
record the strata and that the stratication reasonably
covers Britain. Habitat estimates in general may be
made fromthe calibrated LCM2000 census rather than
the eld sample, taking full account of the hetero-
geneity of a study area, thereby generating more
accurate results. Condence limits for such estimates
are tighter than those from the eld survey. In future, it
is hoped that broad habitat cover estimates might be
proven for local areas. While the 40 strata might
constrain such aims, the local values would be fully
surveyed in LCM2000, and hence more reliable than
equivalent estimates froma eld survey, which had not
visited the locality at all. It may even be possible to
estimate features not directly mapped by LCM2000,
where these can be related quantitatively to other
mapped features, with regional variations accommo-
dated by the environmental strata (e.g. estimating oak
tree cover as regionally weighted proportions of
Broad-leaved woodland).
The exercise emphasises the dilemma facing
producers of user-oriented thematic maps fromremote
sensing. Users had good reasons to demand mapping
of UK broad habitats. The producers had a duty to
inform them that, without compromises on quality, the
target classication was the nearest possible equiva-
lent. Undertaking production in the face of foreseeable
problems was arguably to expend effort, which might
have been invested in areas of likely success. For
example, the classication demanded great efforts in
distinguishing rarer broad habitats, half of them
contributing just 5% of UK cover. From an accuracy
viewpoint, it would have been better to focus on
commonplace classes where small improvements
would have benetted overall quality. However, users
would rightly have questioned the value of an accurate
map which did not show what they wanted.
It is important to recognise LCM2000s inherent
strengths. The combination of the comprehensive
coverage of LCM2000 with the substantially greater
detail available from sample-based eld surveys offers
considerable potential. The generation of calibrated
statistics with condence limits is believed to be novel
in a study of this scale. Another strength is the depth of
information available in the LCM2000 database.
While providing the broad habitat classication,
LCM2000 does not impose the scheme on all. The
class variants, spectral classes and contextual analyses
are recorded for every polygon. It is possible to
regroup classes and recongure contextual analyses;
and for users to do much of this themselves. They can
choose between detail and reliability, aggregating
where their aims allow it, splitting classes where
necessary, adding depth or accuracy by extension of
the contextual analyses.
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 214
In landscape management, policy-needs risk run-
ning ahead of our ability to deliver. Users with
challenging targets, perhaps with budgets, which
render the targets impossible, may ask producers to
adopt solutions, which are currently unworkable.
Remote sensing comes closer to answering users
questions about landscape, land cover and patterns of
land use than any other option for comprehensive
surveys of large areas. However, the development of
viable solutions requires that the producers of
remotely sensed information better understand the
needs of users and users must appreciate what the
technology can and cannot deliver. Knowing the
questions to ask is the rst step towards a workable
solution. For many years, people have reported
communication problems between producers and
users of remotely sensed data. This paper adds the
benets of considerable hindsight to their ongoing
discussions.
Acknowledgements
The paper was written while the lead-author
worked in the Unit for Landscape Modelling in a
post part-funded by the Sir Isaac Newton Trust. The
maps were produced while working for CEH.
References
Barr, C.J., Bunce, R.G.H., Clarke, R.T., Fuller, R.M., Furze, M.T.,
Gillespie, M.K., Groom, G.B., Hallam, C.J., Hornung, M.,
Howard, D.C., Ness, M.J., 1993. Countryside Survey 1990:
Main Report. Department of the Environment, London.
Cherrill, A.J., Lane, A., Fuller, R.M., 1994. The use of Landsat-5
Thematic Mapper imagery in the characterisation of landscape
composition: a case study in northern England. J. Environ.
Manag. 40, 357377.
Cherrill, A.J., McClean, C., Lane, A., Fuller, R.M., 1995. A com-
parison of land cover types recognised in an ecological eld
survey of northern England and in the rst remotely sensed land
cover map of Great Britain. Biol. Conserv. 71, 313323.
Cihlar, J., 2000. Land cover mapping of large areas from satellites:
status and research priorities. Int. J. Remote Sens. 21, 1093
1114.
Comber, A., Fisher, P., Wadsworth, R., 2003. Actor-network theory:
a suitable framework to understand how land cover mapping
projects develop? Land Use Policy 20, 299309.
Comber, A., Fisher, P., Wadsworth, R., 2004a. Assessment of a
semantic statistical approach to detecting land cover change
using inconsistent data sets. Photogramm. Eng. Remote Sens.
70, 931938.
Comber, A., Fisher, P., Wadsworth, R., 2004b. Integrating land-
cover data with different ontologies: identifying change from
inconsistency. Int. J. Geograph. Inf. Sci. 18, 691708.
Congalton, R.G., 1991. A review of assessing the accuracy of
classications of remotely sensed data. Remote Sens. Environ.
37, 3546.
Defries, R.S., Belward, A.S., 2000. Global and regional land cover
characterization from satellite data: an introduction to the
special issue. Int. J. Remote Sens. 21, 3546.
Devereux, B.J., Amable, G., Costa Posada, C., 2004. An efcient
image segmentation algorithm for landscape analysis. Int. J.
Appl. Earth Observ. Geoinf. 6, 4761.
De Wit, A.J.W., Clevers, J.G.P.W., 2004. Efciency and accuracy of
per-eld classication for operational crop mapping. Int. J.
Remote Sens. 25, 40914112.
Edwards Jr., T.C., Moisen, G.G., Cutler, D.R., 1998. Assessing map
accuracy in a remotely sensed, ecoregion-scale cover map.
Remote Sens. Environ. 63, 7383.
Efron, B., Tibshirani, R.J., 1993. An Introduction to the Bootstrap.
Chapman and Hall, London.
European Environment Agency, 1999. CORINE Land Cover: AKey
Database for European Integrated Environmental Assessment.
European Environment Agency, Copenhagen.
Foody, G.M., 2002. Status of land cover classication accuracy
assessment. Remote Sens. Environ. 80, 185201.
Fuller, R.M., 1987. The changing extent and conservation interest of
lowland grasslands in England and Wales: a review of grassland
surveys 193084. Biol. Conserv. 40, 281300.
Fuller, R.M., Groom, G.B., Jones, A.R., 1994a. The land cover map
of Great Britain: an automated classication of Landsat The-
matic Mapper data. Photogramm. Eng. Remote Sens. 60, 553
562.
Fuller, R.M., Groom, G.B., Wallis, S.M., 1994b. The availability of
Landsat TM images for Great Britain. Int. J. Remote Sens. 15,
13571362.
Fuller, R.M., Smith, G.M., Devereux, B.J., 2003. The characterisa-
tion and measurement of land cover change through remote
sensing: problems in operational applications? Int. J. Appl.
Earth Observ. Geoinf. 4 (3), 243253.
Fuller, R.M., Smith, G.M., Thomson, A.G., 2004. Contextual ana-
lyses of remotely sensed images for the operational classication
of land cover in Great Britain. In: de Jong, S., van der Meer, F.
(Eds.), Contextual Methods in Image Processing of Remotely
Sensed Data. Kluwer Academics, Dordrecht.
Fuller, R.M., Smith, G.M., Sanderson, J.M., Hill, R.A., Thomson,
A.G., 2002. The UK land cover map 2000: construction of a
parcel-based vector map from satellite images. Cartograph. J. 39
(1), 1525.
Fuller, R.M., Wyatt, B.K., Barr, C.J., 1998. Countryside survey from
ground and space: different perspectives, complementary
results. J. Environ. Manag. 54, 101126.
Haines-Young, R.H., Barr, C.J., Black, H.I.J., Briggs, D.J., Clarke,
R.T., Cooper, A., Dawson, F.H., Firbank, L.G., Fuller, R.M.,
Furse, M.T., Gillespie, M.K., Hill, R., Hornung, M., Howard,
D.C., McCann, T., Morecroft, M.D., Petit, S., Sier, A.R.J.,
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 215
Smart, S.M., Smith, G.M., Stott, A.P., Stuart, R.C., Watkins,
J.W., 2000. Accounting for Nature: Assessing Habitats in the UK
Countryside. DETR, London.
Henderson, F.M., Lewis, A.J. (Eds.), 1998. Manual of remote
sensing. third ed. Principles and Applications of Imaging Radar,
vol. 2. Wiley, New York.
Homer, C., Huang, C.Q., Yang, L.M., Wylie, B., Coan, M., 2004.
Development of a 2001 National Land-Cover Database for the
United States. Photogramm. Eng. Remote Sens. 70, 829840.
Hornung, M., Bull, K.R., Cresser, M., Ullyett, J., Hall, J.R., Langan,
S., Loveland, P.J., Wilson, M.J., 1995. The sensitivity of surface
waters of Great Britain to acidication predicted fromcatchment
characteristics. Environ. Pollut. 87, 207214.
Jackson, D.L., 2000. JNCC Report No. 307 Guidance on the
Interpretation of the Biodiversity Broad Habitat Classication
(Terrestrial and Freshwater types): Denitions and the Relation-
ships with Other Habitat Classications. Joint Nature Conserva-
tion Committee, Peterborough.
Kershaw, C.D., Fuller, R.M., 1992. Statistical problems in the
discrimination of land cover from satellite images: a case study
in lowland Britain. Int. J. Remote Sens. 13, 30853104.
Prosser, M.V., Wallace H.L., 1999. Countryside Survey 2000:
Quality Assurance Exercise. Unpublished Report to ITE, Merle-
wood.
Ratcliffe, D.A., Thompson, D.B.A., 1988. The British uplands:
their ecological character and international signicance. In:
Usher, M.B., Thompson, D.B.A. (Eds.), Ecological Change
in the Uplands. Blackwell Scientic Publications, Oxford,
pp. 936.
Reid, E., Quarmby, N., 1997. Blanket coverage: mapping and
assessing Scotlands bogs. Mapp. Awareness 11, 3437.
Schowengerdt, R.A., 1997. Models and Methods for Image Proces-
sing. Academic Press, London.
Smith, G.M., Fuller, R.M., 2000. Land cover map 2000: more than
just an update. Soc. Bull. 33 (2), 1316.
Smith, G.M., Fuller, R.M., 2002. Land cover map 2000 and meta
data at the land parcel level. In: Foody, G.M., Atkinson, P.M.
(Eds.), Uncertainty in Remote Sensing and GIS. Wiley and
Sons, London.
Smith, G.M., Fuller, R.M., Hoffmann, A., Wicks, T., 2000. Parcel-
based approaches to the analysis of remotely sensed data. In:
Proceedings of the Remote Sensing Society Conference, Adding
Value to Remotely Sensed Data. Remote Sensing Society,
Nottingham CD-ROM.
Smith, G.M., Fuller, R.M., Amable, G., Costa, C., Devereux, B.J.,
1997. CLEVER-Mapping: an implementation of a per-parcel
classication procedure within an integrated GIS environment.
In: Proceedings of the Remote Sensing Society Conference,
Observations and Interactions: RSS97, Remote Sensing Society,
University of Nottingham, pp. 2126.
Thunnissen, H.A.M., De Wit, A.J.W., 2000. The national land-cover
database of the Netherlands. In: Geoinformation for All; XIX
Congress of the International Society for Photogrammetry and
Remote Sensing (Lemmer: GITC). pp. 223230.
UK Biodiversity Steering Group, 1995. Biodiversity: The UK
Steering Group Report, vols. 1 and 2, HMSO, London.
UNEP, 1992. Convention on Biological Diversity. United Nations
Environment Programme, Nairobi.
Van Oort, P.A.J., Bregdt, A.K., De Bruin, S., De Wit, A.J.W., Stein,
A., 2004. Spatial variability in classication accuracy of agri-
cultural crops in the Dutch national land-cover database. Int. J.
Geograph. Inf. Sci. 18, 611626.
Wyatt, B.K., Greatorex-Davies, N.G., Bunce, R.G.H., Fuller, R.M.,
Hill, M.O., 1994. Comparison of Land Cover Denitions.
Countryside 1990 Series: Volume 3. Department of the Envir-
onment, London.
R.M. Fuller et al. / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 202216 216