types (Loveland et al., 1991). Vegetation indices derived
from remotely sensed data have been found to be
sensitive to the presence and condition of green
vegetation. The Normalized Difference Vegetation Index
(NDVI) is one of the commonly used vegetation indices
(Goward et al., 1985). The U.S. Geological Survey, EROS
Data Center has been compiling bi-weekly NDVI
composites that identify peak vegetation growing
conditions in each two-week period since 1990. One
composite from each month of the growing season in
1991 was selected as input to the model for vegetation
discrimination. The eight bi-weekly composites are listed
in Table 1. The June image of the study area is shown in
Figure 1.
Table 1. List of Bi-Weekly NDVI Images
Period Month Dates
1 March 15-28
2 April 12-25
3 May 10-23
4 June 07-20
5 July 05-18
6 August 02-15
7 September 13-26
8 October 11-24
The State Soil Geographic (STATSGO) data set is a digital
general soil association map developed by the National
Cooperative Soil Survey, USDA. It depicts information
about soil features such as water capacity, salinity, layer
depth, and organic matter. The STATSGO data are
1:250,000 scale statewide coverages generalized from
more detailed soil survey maps at scales ranging from
1:12,000 to 1:63,360 (USDA, 1994). The purpose of using
the STATSGO data in the model is to reduce data volume
of remotely sensed images by summarizing ecological
parameters (e.g. vegetation indices) within map units, and
to use the polygon neighborhood relationship of vector
maps to maintain the contiguity of regions.
METHODOLOGY
The regionalization model generates regions from
polygonal map units according to their similarity in biotic
and abiotic parameters. ^ Regions remain spatially
contiguous since only neighboring units can be assigned
into one cluster in the aggregating process. The base
map units are usually low level ecosystems that exhibit
relative homogeneity of ecology properties of interest.
The parameters are ecological features used to define
ecoregions. For the vegetation zoning of Nebraska, the
STATSGO map units were used as base units while
multitemporal NDVI images were used as parameters for
similarity computation. The regionalization model was
written in C programming codes. The output maps could
be displayed in GIS or remote sensing software packages.
Data Pre-processing
The eight NDVI images of the study area were
transformed into Albers Conical Equal Area Projection to
match the projection of the STATSGO map. Generalizing
parameter values within base map units was the first step
in the modelling process. Each of the eight NDVI images
was overlaid with the rasterized STATSGO map and its
pixel values were averaged within STATSGO map units.
One of the generalized images is shown in Figure 2. After
this process, each map unit was associated with a set of
generalized NDVI values. The neighboring relationships
between polygons (units) were extracted from GIS
topological files. This data set was used to control the
contiguity of ecoregions.
Region Partition
The regionalization model generated regions by
aggregating base units through a hierarchical clustering
procedure. Two criteria were used to control cluster
mergers. One was the similarity between map units or
clusters. The similarity index (SI) was calculated from
Euclidean distance using the following formula:
SÉ. zx Y (Xu Xi)” kz1,2,3,...8
where, X; is the averaged NDVI value of band k for map
unit i, and
X is the averaged NDVI value of band k for
map unit j.
For every two polygons, the smaller the similarity index,
the closer their ecological characteristics. The other
criterion was the neighborhood relationship between
polygons. Two units can be assigned to one cluster only
if their distance is small and they are spatially next to each
other. The aggregating process iteratively searched for the
pair of polygons with the smallest SI, merging them into
one cluster, recalculating the area-weighted NDVI values
for the new unit, and calculating new distance from this
unit to its neighbors. Consequently, clusters grow larger
and larger until they are large enough to be final regions.
Since only spatially connected units were merged, the
regions remained contiguous while they were growing.
For Nebraska, the regional clustering started from 2,024
units (i.e., the total number of STATSGO map units in the
state). The last 51 mergers are shown in the dendrogram
of Figure 3. Each of the 51 clusters comprises more than
10 original units. Twelve regions were formed at distance
level around 0.5, two of them were too small in area and
dissolved into the surrounding regions. The dendrogram
was broken down at this level because of two reasons.
One was that the areal extent of regions at this stage are
close to that of other ecoregion schemes for this area, the
other was that several large regions (2, 3, and 4) merged
at this level almost simultaneously.
1002
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
