Multiscale Hierarchical Classification of Wetland Environments
Using SAR Data
Michael R. Ricard and Melba M. Crawford
Center for Space Research, University of Texas at Austin, 3925 W. Braker Ln.,
Suite 200, Austin, TX 78759-5321
Ph: (512) 471-5573; Fax: (512) 471-3570; Email:
crawford@csr.utexas.edu
Abstract - A hierarchical, multiresolution classification approach based
on an ensemble of neural networks in scale has been implemented and its
capability investigated for discriminating between wetland vegetation classes
in fully polarimetric, multifrequency synthetic aperture radar (SAR) data.
Problems with poorly separated spectral signatures between vegetation types
coupled with speckle degrade the quality of standard pixel-based approaches.
Because classes differ in spatial extent and are often very irregular in shape,
standard region based approaches do not perform well either. The approach used
in this study involves sequential separation of land cover classes within the
multiscale framework. The algorithm was applied to a coastal salt marsh on the
Texas coast where AIRSAR data were acquired in 1996. Classification results
were compared to those obtained via a standard Maximum Likelihood procedure and
a simple feed-forward neural network formulation.
INTRODUCTION
Classification of land cover is one of the primary objectives in the analysis
of remotely sensed data. In synthetic aperture radar (SAR) data, speckle,
variation in scattering coefficient with incidence angle, and differences in
statistical separability between classes all affect the accuracy of pixel-based
classifiers. By compensating for these sources of variation in backscatter
statistics, a more accurate land cover map can be created.
Filtering techniques are commonly applied to the input data for the reduction
of speckle in SAR imagery. While these filters improve the appearance of the
image for visual interpretation and result in smoother classification maps,
statistical variability in measured backscatter caused by speckle still remains
at the pixel level. To compensate for this variability, neighborhood, or
contextual, information is often incorporated into the classifier methodology.
One such method utilizes multiresolution, or multiscale, techniques which model
the statistics of a given pixel and class from coarse to fine scale
resolutions. This information can then be used to perform the classification
of remotely sensed imagery, as in [1].
When discriminating between land cover types which have groupings with similar
signatures, it is often beneficial to first separate the data into general
groupings rather than specific classes, as well as to remove those classes
which are easily separable from the rest. The complexity of the classifier
tends to be reduced, both in terms of training and decision making, since the
classifier does not need to distinguish between all classes simultaneously [2].
In addition, in analysis of SAR data from coastal environments, variation due
to changes in angle of incidence may cause otherwise separable classes, namely
water and land, to overlap statistically. A hierarchical classifier approach
can also be used to resolve these conflicts in class separability because the
data can be class-corrected at any specified level of the classifier using
models for the variation in backscatter with angle [3]. Examples of
hierarchical classifiers applied to remotely sensed imagery are contained in
[4] and [5].
The following sections contain descriptions of the classification of wetland
environments using multifrequency, fully polarimetric airborne SAR data.
Results are shown for a multiscale and a hierarchical classifier based on
feed-forward neural networks. A classifier which combines these methods is
presented, and results of the classifiers are compared.
TEST SITE AND DATA DESCRIPTION
Bolivar Peninsula is part of the low relief barrier islands of the Texas coast
located at the mouth of Galveston Bay. The test site chosen for this study is
located at a washover fan on southern Bolivar Peninsula. This area can
generally be grouped into wetland and upland communities, with the marsh area
characterized in terms of sub-environments defined by wetland maps [6]. For
classification purposes, the land cover types which occur in these environments
include 11 classes comprised of water, wetlands (low proximal marsh, high
proximal marsh, high distal marsh, and brackish marsh), and uplands (trees,
agriculture, urban, dunes/flats, and grass/general uplands).
For the purpose of mapping this wetland vegetation, NASA/JPL airborne SAR
(AIRSAR) data were collected on June 28, 1996 over Bolivar Peninsula on the
Texas coast. Fully polarimetric radar data (20 MHz) were acquired at C, L, and
P-bands. The data were ground range projected resulting in pixels with 10m
spatial resolution. Classifier inputs consisted of C-band and L-band linear
polarizations (HH, VV, HV), circular polarizations (RR, RL), and phase
differences (HHVV). As a pre-processing stage, the AIRSAR data were speckle
filtered using Lee's single-frequency polarimetric SAR speckle filter [7].
CLASSIFICATION METHODOLOGY
Neural Network Architecture
A single hidden layer feed-forward neural network, multilayer perceptron
(MLP), was chosen as the basis for the various classifier architectures
presented. The initial network consisted of 14 input units, 25 hidden units,
and 11 output units. Each network was trained using the scaled conjugate
gradient supervised learning algorithm, on approximately 400 training samples
per class, with the targets specified using the 1-of-c encoding scheme to
generate outputs which estimated the posterior probability for each class [8].
To preserve classifier generalization, a validation set was used to prevent
over-training of each network. To remove unnecessary nodes, non-contributing
input and hidden units of each network were pruned from the network based on a
statistical measure of a unit's contribution to the network's behavior [9].
This resulted in different numbers of input and hidden units for each of the
networks, but did not significantly affect accuracy.
Multiscale Classifier Architecture
A multiresoluton framework was employed to account for the presence of speckle
in the imagery and incorporate neighborhood information in the classifier. A
block diagram of the multiscale classifier architecture is depicted in Figure
1.
Figure 1. Multiscale Classifier Architecture.
First, the 14 input channels of the original 10m spatial resolution data were
decimated using a separable 2-d hamming window [1]. This reduced the spatial
resolution by nearly a factor of 2 (1.81) with the additional benefit of
sidelobe reduction for high intensity scatterers which are prevalent in urban
areas. The resulting data sets had 18m and 32m spatial resolution
respectively.
The three multiscale data sets were inputs to individual MLP classifiers. The
networks were each trained as above producing posterior probability estimates
for the classes at each resolution. To combine the results from the three
scale classifiers, a linear opinion pool was used, as described in [2], [10],
and [11]. A simple average was performed on the scale classifier outputs
producing an ensemble multiscale probability.
Hierarchical Classifier Architecture
A hierarchical classifier architecture depicted in Figure 2 was devised to
deal with variable separability between classes, and the class dependent
variation in backscatter associated with changes in incidence angle
Figure 2. Hierarchical Classifier Architecture.
The classification process was comprised of three levels. At successive levels
in the hierarchical structure, the inter- class separability between classes
decreases, but is approximately uniform within the level. This allows the
classifier at any level to focus its decision making process on a small number
of similarly separated classes instead of attempting to distinguish many
classes with varying degrees of separability.
Different inputs can be used at each level of the hierarchy. For the SAR data
in this study, the greatest variation in radar backscatter due to changes in
angle of incidence occurs over the open water. To correct for this variation,
water pixels were isolated, and linear corrections, based upon fits to a model
derived from [3], were applied at each frequency and polarization. The "water
corrected" data were inputs to the Level-1 and Level-2a classifiers, while the
"original" data was an input to the Level-2b, Level-3a, and Level-3b
classifiers.
Multiscale Hierarchical Classifier
As the last stage, a multiscale hierarchical classifier was implemented
combining the two methods described previously. To that end, the fine scale
classifier at each level of the hierarchy was replaced by its multiscale
counterpart.
CLASSIFICATION RESULTS
Airborne SAR data acquired over a wetland environment along the south Texas
coast were classified first via a standard Maximum Likelihood (ML) classifier
and a single MLP neural network classifier. The multiscale classifier was then
applied to improve accuracy through the suppression of speckle, while the
hierarchical classifier was utilized to correct for the variation in
backscatter of the water class and separate the classification process into
levels of comparably separable classes. Finally, the combined multiscale
hierarchical classifier was applied to the SAR data. The test set
classification accuracy for each of these methods is shown in Table 1, with the
best results obtained via the multiscale and multiscale hierarchical
procedures. In examination of the classified imagery, both hierarchical
classifiers preserved fine scale water features better than the other
classifiers.
CONCLUSION
SAR data issues which influence the classification process of coastal
environments, namely the presence of speckle, variation in backscatter with
angle of incidence, and differences in statistical separability of classes,
were the focus of this paper. To compensate for these issues, multiscale and
hierarchical classifier architectures were devised. Results obtained from
application of these classification methods to SAR data acquired over a wetland
environment are superior to both standard statistical and neural network
methods, and support the usefulness in incorporating multiscale and
hierarchical architectures in the classification process.
REFERENCES
- W.W. Irving, L.M. Novak, and A.S. Willsky, "A Multiresolution Approach to
Discrimination in SAR Imagery", IEEE Transactions on Aerospace and
Electronic Systems, vol. 33, no. 4, pp. 1157-1169, October 1997.
- A.J.C. Sharkey, "On Combining Artificial Neural Nets", Connection
Science, 8(3/4), 1996.
- F.T. Ulaby, R.K. Moore, and A.K. Fung, Microwave Remote Sensing: Active
and Passive, Volume II: Radar Remote Sensing and Surface Scattering and
Emission Theory, Boston: Artech House, 1982.
- M.M. Crawford and M.R. Ricard, "Hierarchical Classification of SAR Data
Using a Markov Random Field Model", in press.
- M.C. Dobson, L.E. Pierce, and F.T. Ulaby, "Knowledge-Based Land-Cover
Classification Using ERS-1/JERS-1 SAR Composites", IEEE Transactions on
Geoscience and Remote Sensing, vol. 34, no. 1, pp. 83-99, January 1996.
- W.A. White, T.R. Calnan, R.A. Morton, R.S. Kimble, T.G. Littleton, J.H.
McGowen, H.S. Nance, and K.E. Schmedes, Submerged Lands of Texas,
Galveston-Houston Area, Bureau of Economic Geology, University of Texas at
Austin, 1985.
- J.S. Lee, M.R. Grunes, and G. De Grandi, "Polarimetric SAR Speckle
Filtering and Its Impact on Classification", In Proceedings of International
Geoscience and Remote Sensing Symposium, vol. II, pp. 1038-1040, August
1997.
- C.M. Bishop, Neural Networks for Pattern Recognition, Oxford:
Clarendon Press, 1995.
- Stuttgart Neural Network Simulator (SNNS) User Manual, Version 4.1,
University of Stuttgart, Institute for Parallel and Distributed High
Performance Systems (IPVR), 1995.
- M.R. Ricard, A.L. Neuenschwander, M.M. Crawford, and J.C. Gibeaut,
"Multisensor Classification of Wetland Environments Using Airborne
Multispectral and SAR Data", In Proceedings of International Geoscience and
Remote Sensing Symposium, vol. II, pp. 667-669, August 1997.
- J.A. Benediktsson, J.R. Sveinsson, and P.H. Swain, "Hybrid Consensus
Theoretic Classification", In Proceedings of International Geoscience and
Remote Sensing Symposium, vol. III, pp. 1848-1850, 1996.
Table 1. Classification Accuracy for Test Sets.
|
Probability
of Correct Classification
|
|
Class
|
|
Classifier
|
1
|
2
|
3
|
4
|
5
|
6
|
7
|
8
|
9
|
10
|
11
|
Overall
|
ML
|
96.5
|
98.5
|
99.5
|
94.3
|
75.7
|
91.3
|
61.9
|
98.5
|
97.3
|
49.8
|
73.3
|
85.1
|
MLP
|
98.0
|
97.3
|
96.0
|
94.1
|
98.8
|
88.4
|
79.5
|
94.8
|
90.6
|
46.8
|
93.3
|
88.9
|
Multiscale
|
100.0
|
99.5
|
100.0
|
98.0
|
99.5
|
94.6
|
89.6
|
99.3
|
98.5
|
69.6
|
98.0
|
95.1
|
Hierarchical
|
99.3
|
99.0
|
94.8
|
96.0
|
91.8
|
91.3
|
81.9
|
90.1
|
94.6
|
57.9
|
94.6
|
90.1
|
Multiscale
Hierarchical
|
100.0
|
99.8
|
98.3
|
99.8
|
99.3
|
94.8
|
90.3
|
96.0
|
98.0
|
71.8
|
97.5
|
95.1
|