ds
nds
r and
sion,
high
ction
ctual
ys in
mote
>d on
ation
à and
oring
ze of
patial
vhich
> and
nsing
ns in
rious
nutes
titute
lands
1994
The
land,
nsive
t and
1 cell
ut 10
ering
Land
DLO,
matic
of the
available geographic land cover information for
pan-Europe. Four data bases were used: the RIVM 10
Minutes pan-European Land Use data base, the
Bartholomew Euromaps CD-ROM, the ESA Digital Forest
Map (ESA, 1992) and the Digital Chart of the World,
originating from the US Defence Mapping Agency. The
classes of the various data bases were either implemented
directly or used in (conditional) combinations in order to
obtain an optimal reference map of pan-European land
cover in terms of accuracy and nomenclature. In this way
18 different classes of land cover are distinguished.
Towards European land cover mapping using remote
sensing there are very interesting approaches, e.g. the visual
interpretation of Landsat-TM and SPOT-XS hard copies in
the CORINE land cover project (CORINE, 1993), taking
texture, pattern and context into account. The scale of this
land cover data base is 1 : 100,000. However, CORINE
does not produce an agricultural, but an ecological legend.
Landsat-TM and SPOT data have, contrary to
meteorological satellite data, a high spatial resolution and
therefore it will take a very long time to map the whole of
e.g. Europe or Russia. Besides, a frequent update is
impossible. Moreover, the total processing time and the
expenses would be unacceptable for monitoring purposes.
So, to monitor the land cover for an extended area like
Europe or Russia, low resolution satellite data such as
NOAA-AVHRR data seem to be more suitable. In a pilot
study Mücher et al. (1994) used NOAA-AVHRR data for
deriving as main land cover classes at continental scale:
arable land, grassland, forest, built-up areas and water.
A major initiative has been the 1 km land cover project of
the International Geosphere Biosphere Programme's Data
and Information System (IGBP-DIS). This had the goal of
collecting, archiving and processing daily NOAA-AVHRR
data for all terrestrial surfaces and then deriving land cover
data sets from this archive. As a result, a 1 km global land
cover product, called DISCover, has been created (IGBP-
DIS, 1996). The applied land cover classification scheme
consists of 17 classes. Because the method is based on
unsupervised classification of monthly NDVI composites
per continent, the result is more related to an agro-
climatologic zonation.
In this paper methods are described to monitor agricultural
land cover applying multi-sensor remote sensing
techniques. In this study the land cover classes, which can
be derived at different scale levels, will be identified. Main
goal is the improvement of agricultural land use monitoring
in Russia. A methodology must be developed, based on
remote sensing, with a much higher accuracy than
traditional methods.
This project is performed in the context of a programme for
Russian-Dutch co-operation in the field of agricultural
research, aiming at exchange of knowledge and experience
concerning remote sensing methodologies and satellite
systems, including also training activities.
level 1 NOAA time series
National & Regional
‚Statistics
Biophysical
Ancillary Data
Major Land Cover Classes Major Landscape Zones
Agricultural Regions
level 2 MSU-SK
Data
L Cra lia A
Delineation of
Agricultural Areas
Grassland/Cropland
level 3 SPOT, MSU-E,
Landsat-TM Data 0
"RTT SUE.
level 4 | TK-350,
| KVR-1000 Data Fem
L———————
Crop Types
Specific Small Land Cover Types
and
Local Phenomena
Figure 1. Schematic overview of the methodology.
2. METHODOLOGY
Figure ] presents a schematic overview of the multiscale
methodology proposed in this study. First of all, ancillary
data on soil, altitude, climate, etcetera, in combination with
statistical data are used for deriving the major landscape
zones. Then, a multi-temporal selection of coarse resolution
images (e.g., NOAA-AVHRR) is made depending on the
sowing and harvesting schemes of a region. These images
are then classified into the major land cover classes (level
1) using the stratification of the whole area into the major
landscape zones (Addink, 1997). This yields the major
agricultural regions.
The resulting classes 'arable land' and 'grassland' are
subsequently checked, delineated more accurately and, if
possible, further differentiated using medium resolution
satellite data (around 200 m spatial resolution; level 2).
Such data (e.g; MSU-SK data) will also facilitate the
comparison of derived land cover classes from coarse
resolution and high resolution images.
As a next step, the pixels classified as arable land or
grassland, the so-called agricultural pixels, are classified
according to a more detailed legend (level 3) using high
resolution images (e.g., SPOT, MSU-E, Landsat-TM). The
contents of this legend depend naturally on the crops.
present in the region. Since it is unfeasible to cover the
entire Russian territory with high resolution images a few
test sites are studied in more detail. The selection criteria
for these sites are the percentage of agricultural pixels, the
availability of ground truth data and the available imagery.
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 97
