International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
50
The problems with mixed pixels can also be alleviated by
applying subpixel analysis as a preprocessing step, before
segmentation. Spectral and spatial subpixel analysis may be
distinguished.
Artificial neural networks, as an example for nonparametric
classification, do not require normally distributed features and
therefore work well with all types of features (Bischof et al.,
1992). The drawback of artificial neural networks is the large
quantity of training data needed: The statistical distribution of
features for each class has to be derived from the training data
only, as there is no a-priori assumption on the type of feature
distribution.
“Hard” decision tree classification and its “soft” versions based
on probability theory, evidence theory and possibility theory,
are well suited for using generic a-priori-knowledge on the
feature characteristics in the different classes. Relaxation
labelling (see e.g. Pal and Pal, 1993) is especially suited for
exploiting relations between image objects and thus is also of
interest in landcover classification.
For the work reported here, both conventional maximum
likelihood classification and decision tree classification based
on generic knowledge on the spectral reflectance of landcover
categories have been used.
Generic classification rules can be employed and, as a
consequence, the costly provision of image-specific training
data can be avoided, if the images are radiometrically calibrated.
Radiometric calibration transforms pixel values into terrain
reflectance values, accounting for changes in ground surface
irradiance due to varying sun elevation angle and terrain slope,
for absorption and scattering processes in the atmosphere as
well as for the influence of sensor parameters. The main
difficulty of radiometric calibration is the procurement of
atmospheric parameters necessary for the quantitative
simulation of absorption and scattering mechanisms in the
atmosphere. It is possible to obtain estimates of the required
atmospheric parameters from the images themselves. This
approach can be termed radiometric self-calibration
(Steinwendner and Schneider, 1999).
Different landcover classes might have similar or even identical
spectral and textural signatures, so that they cannot be identified
from the remotely sensed images alone. Reference information,
e.g. from digital terrain models or from other GIS data sets of
various types may be used in these cases together with the
remotely sensed data.
The transformation of landcover information to landuse
information, which is beyond the scope of this contribution, will
in any case heavily rely on context information (spatial relations
between objects), and on external information.
4. STRATEGY FOR INTERRELATED
SEGMENTATION AND CLASSIFICATION
Figure 7 illustrates the workflow of the interrelated
segmentation and classification method for automated landcover
mapping. The first step is the radiometric calibration of the
input image. A self-calibrating procedure is employed, making
no use of external calibration data, and working with
information from the input image only. The main idea here is to
estimate atmospheric path radiance from the lowest pixel values
in the individual spectral bands, derive aerosol concentration
parameters from this (based on the assumption of the presence
of typical atmospheric models for the gaseous components and
of typical aerosol types and distributions), and to compute the
atmospheric radiation quantities needed for the calibration with
the 6S code (Tanre et al., 1990, Vermote et al., 1997).
Pixelwise maximum likelihood classification is applied to the
calibrated image. Generic probability density distribution
parameters (mean values and covariance matrices for each
category) can be used for this task. Only categories that can be
spectrally separable should be chosen at this stage. Landcover
categories with bad spectral separability should be combined in
joint spectral categories to be separated later. For each pixel, not
only the category label with the maximum likelihood, but also
the Mahalanobis distances to each category mean are stored. A
seed image (image of seed candidates) is produced from the
pixelwise classification result by selecting reliably classified
pixels on the basis of the Mahalanobis distance. Seed
candidates are not allowed in high gradient areas of the image,
in order to avoid mixed-pixel seeds.
The calibrated image is segmented by region growing, starting
from seed pixels as specified in the seed image. The landcover
category (or at least the joint spectral category) is known for
every seed pixel. Therefore, class-specific region growing
parameters (thresholds, absolute spectral intervals) can be
employed. This first segmentation is performed with rather
restrictive threshold values (representing the initial
segmentation parameters, see Fig. 7). As a consequence, only
very homogeneous segments will emerge. A large fraction of
pixels will not be assigned to segments. Segments of classes
which are typically rather inhomogeneous (e.g. settlement areas)
will not be formed at all in this processing step.
A test for mixed pixels follows, as described in chapter 2.2
above. Subpixel candidates are assigned to the spectrally nearest
segment. In cases where pure spectral information is insufficient
for the classification of the final landcover categories, it may
now be refined using shape parameters and context information
E.g., grassland and agricultural fields may now be differentiated
on the basis of segment shape. A considerable fraction of mixed
pixels has been eliminated (assigned to segments) at this stage.
The segmentation parameters (Mahalanobis distance constraints
for seed pixels, thresholds for region growing) may now be
redefined, allowing for the formation of more inhomogeneous
segments, e.g. in settlement areas. Segmentation and refined
classification are repeated, until all pixels have been assigned to
segments and classified.
