Determining detailed vegetation characteristics to classify arid
rangelands often presents unique problems due to the high reflectance of
the soil background, a mixture of green and senescent grasses, and the
prevalence of shrubs in grasslands. These components can make it
difficult to determine the proportion of grass cover. On the Jornada
Experimental Range (JER), operated by the U.S. Department of Agriculture
Agricultural Research Service near Las Cruces, N.M., ongoing research is
aimed at determining the relationship between ground-based observations
and remotely sensed data.
The goal of a recent study was to develop a detailed vegetation
classification of a 1,200-hectare pasture to determine the extent of
grassland and identify locations, extent and percent cover values for
several grass species. Specific objectives were to develop and evaluate
near-Earth photography for ground truthing a QuickBird satellite image
from DigitalGlobe (www.digitalglobe.com); conduct a multiscale analysis,
including ground sampling, near-Earth photography and satellite imagery;
and combine object-oriented classification with classification and
regression trees to analyze the satellite image.
Research Methods
The research team conducted extensive field sampling (325 plots) by
photographing ground vegetation from a height of 2.8 meters.
Thresholding techniques determined percent vegetation cover and
percent bare soil. Fifty plots were chosen for detailed ground
sampling for comparison with the results from the image analysis.
The QuickBird image was analyzed with an object-oriented approach,
using eCognition image-classification software from Definiens
Imaging (www.definiens-imaging.com).
The first step involved segmenting the image based on scale, color
(spectral information) and shape to identify object primitives based
on the chosen parameters. Classification is then performed using
those objects rather than single pixels. The classification is based
on fuzzy logic theory combined with user-defined rules. The
segmentation was performed at two different scales to construct a
hierarchical network of image objects representing the image
information in different spatial resolutions simultaneously. This
allowed the research team to differentiate individual shrubs on a
lower level and delineate broader landscape classes on a higher
level. After the shrubs were classified, they were “removed” from
the image so the segmentation at the higher level didn’t include
their spectral values. This approach allowed the researchers to
determine the shrub-interspace vegetation.
For each image object containing the field plot, several spectral,
spatial and texture characteristics were extracted from the image.
Ancillary information included soils, elevation, aspect and slope
layers. The data were analyzed using classification and regression trees
to determine correlations between features of the segmented image
objects and the measured field plot parameters. A decision tree is a
tool for determining which features are most appropriate for predicting
a particular class based on ground-plot information. With the help of a
decision tree, one can quickly sift through numerous features associated
with the image objects and select the best ones. Because eCognition
offers hundreds of spectral, spatial and textural features, a decision
tree is a useful tool for reducing the large number of input variables.
Favorable Results
Results for percent vegetation cover and bare soil calculated from the
near-Earth photography showed close correlation to the ground-based
sampling and proved to be a faster assessment tool than ground sampling.
One disadvantage was the occurrence of shadow in the photos, which could
be eliminated by using shading for the plot.
The object-oriented classification of the QuickBird image worked
favorably, because shrubs could be classified separately at a finer
scale while the shrub-interspace vegetation could be analyzed at a
coarser scale. This allowed the research team to get a reliable estimate
of grass cover and shrub density in the pasture. The rule base derived
from the decision tree proved to be successful at differentiating
between the dominant grass species as well as defining several classes
of percent grass cover. Future research will include refining the
predictive ability of the decision tree and determining the possibility
of applying this model to other locations and/or to other scales.
.
