rks
twork
(good
show
ihood
cation
roved
ntory
ok was
> over
r crop
(MD)
crop
s are
Sul
ry to
non-
h to
es of
ssing
hown
stical
nulti-
high
er of
on ef
have
igher
from
borne
efore,
N for
yy of
crop
oach.
ing a
data,
SAR
ation,
* to evaluate ANN (non-statistical) as a post-segmentation
classifier in comparison to a minimum distance (MD,
non-parametric) classifier.
STUDY AREA AND DATA DESCRIPTION
The study area is situated in an agricultural area in Oxford
County, southern Ontario, Canada. This site has been used
previously for intensive study of the relationships between
radar data and agricultural parameters. The major field crops
include corn, soybeans, winter wheat, barley/oats, alfalfa
and pasture.
Three dates of early- and mid-season ERS-1 C-VV SAR data
were acquired during 1992 growing season (June 15, July 24
and August 5). July 24 SAR data were acquired in ascending
mode, while others were acquired in descending mode. One
date of Landsat TM data were also acquired on August 6,
1992. Detailed field information was collected at the time of
the overpasses and was input to a geographic information
system to aid in developing and understanding the
classifications.
METHODOLOGY
In the analyses presented in this paper, single-date SAR
data, multi-temporal SAR data, and combinations of SAR and
TM data were classified. In all cases, a per-field
classification approach is adopted since this conforms to
conventional mapping strategies and has been widely used
in radar remote sensing as a means of reducing the effect of
speckle (Foody et al., 1994; Ban et al., 1995). ANN was
used in post-segmentation classifications and was compared
to a MD classifier. MLC was also performed for comparison
purposes.
Pre-preparation
The raw signal SAR data were processed by Atlantis
Processor at the Canadian Center for Remote Sensing and
geometrically corrected to field boundaries (Universal
Transverse Mercator -UTM projection). The geocoded field-
boundary file for the study area was digitized from SPOT
imagery in a GIS and then imported into an image
processing system. To eliminate the effects of field-
boundary pixels and minor image registration errors on crop
discrimination, a 5-pixel buffer was applied to the field
boundaries. This procedure is similar to that used by Ban et
al. (1995).
Calibration and Validation Blocks Selection
The major crops classified in this study were winter wheat,
corn, soybeans, barley/oats, alfalfa and pasture/cut-hay-
alfalfa. Due to the differences in growing stages and ground
cover density, corn and soybeans were further divided into
two classes, good growth or poor growth. For each crop,
pixel sample blocks were randomly extracted within
representative fields in order to calibrate the minimum-
distance-to-means classifier and to train the artificial neural
network.
To assess the accuracy of the classifications, validation
pixels, independent from the calibration pixels, were
randomly selected for each crop. Fields that exhibited
anomalies, such as spectral reflectance/backscatter that
deviated significantly from the norm of a particular class,
49
were excluded from both the calibration and validation
samples. These anomalies usually resulted from weeds
infection, crop management and/or soil drainage
characteristics. The calibration and validation blocks
selection were based on the crop information, i.e. crop
growth stage, ground cover, height, row direction, etc. in a
PAMAP GIS.
Calibration and validation pixels were extracted from
different fields, a requirement for the field approach where a
field was defined as a homogeneous area and all pixels were
assigned the mean value of the field. This reduced the
number of fields that could be used for calibration and
validation, so calibration had to be restricted to fewer fields.
Per-pixel Classification
In order to assess the effectiveness of the non-parametric and
non-statistical approaches, a comparison to a MLC was
required. A number of classifications for SAR, TM and their
combinations using MLC were performed.
Per-field Classification
Since a field only grows a single crop in Canada, it is
desirable to use a per-field classification. Also a per-field
approach reduces the SAR speckle effects, as discussed
earlier.
Segmentation. A per-field classifier permits
segmentation of the ERS-1 SAR data into homogeneous
fields using field boundaries. A unique grey level was
assigned as a label to each output polygon of the field-
boundary file which was then input for the homogeneous
classifier as a theme channel. The homogeneous classifier
defined the homogeneous segments of interest. There were
two values that could be assigned to segments, namely the
mean and the mode. Only the mean was tested in this study.
The pixel values in each field were replaced with the mean
value for that field.
Post-segmentation Classification. Two post-
segmentation classifiers, MD and ANN, were investigated.
MD is the minimum Euclidean distance classifier. It assigns
each pixel to the class which has the minimum distance
between the pixel value and the class mean. In situations
where MLC's multivariate normal distribution assumption
does not hold, MD may perform better than MLC, because
MD does not require making assumptions. In this study,
EASI/PACE software, MINDIS was used (PCI, 1994).
ANN classifiers provide an emerging paradigm for pattern
recognition implementation that involves large
interconnected networks of relatively simple and typical
non-linear units (i.e., neural nets) (Schalkoff, 1992; Foody
et al., 1995). A neural network consists of interconnected
processing elements called units ("nodes" or "neurons").
These are organized in two or more layers. There is an input
layer of units which are activated by the input image data.
The output layer of units represents the output classes to
train for. In between, there is usually one or more hidden
layers of units. An Artificial Neuron Computational
Structure is shown in Figure 1.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996