Peer-Reviewed Articles
1149 Bayesian Soft Classification for Sub-Pixel
Analysis: A Critical Evaluation
J.R. Eastman and R.M. Laney
Abstract Download
Full Article
Soft classifiers defer the decision about the class membership
of a pixel in favor of an expression of the degree of membership it exhibits
in each of the land-cover classes under consideration. The reasons
for using a soft classifier include the examination of classification
uncertainty, but are most commonly directed to the potential of uncovering
the proportional constituents of mixed pixels--a process called sub-pixel
classification. In this study we examine the assumptions and procedures
of a commonly cited Bayesian soft-classification procedure for sub-pixel
classification, and test its ability to uncover mixture proportions.
The procedure involves the use of mixed-cover training sites to estimate
the underlying class signatures through the development of fuzzy mean
reflectances and covariance matrices. These are then used to evaluate
the Bayesian a posteriori probability of belonging to each land-cover
class. Using an artificial data set, it was found that this Bayesian
soft-classification procedure is unable to uncover constituent class
proportions unless substantial overlap exists in the distributions
of parent classes. It was found that the use of fuzzy training sites
improves the accuracy of this procedure, but not because of any special
insights it offers into the underlying distributions, but rather, because
of its tendency to increase the degree of overlap between parent distributions.
1155 The Effect of Training Strategies on Supervised
Classification at Different Spatial Resolutions
DongMei Chen and Douglas Stow
Abstract Download
Full Article
Three different training strategies often used for supervised classification--single
pixel, seed, and block or polygon training--are compared in this paper. The
range parameter of semi-variograms obtained from sample image subsets of each
land-use;shland-cover class was used to measure the autocorrelation level during
training set selection. Eight training sets with different sizes were generated
and then applied to image subsets with three multispectral bands and variance
texture images in the classification of six land-use classes. The classification
results using these training sets were compared at five resolution levels and
were based on six Color Infrared Digital Orthophoto Quarter Quadrangle (DOQQ)
subsets of different urban land types in urban and rural fringe areas of the
San Diego metropolitan area. The performance of different training strategies
is shown to be influenced by the training size, the image resolution, and the
degree of autocorrelation inherent within each class. Training approaches had
more impact on classification results at fine resolution levels than at coarse
resolutions. For spectrally homogeneous classes, a spatially independent, single-pixel
training approach is preferred. But for spatially heterogeneous classes, small
block training has the advantage of readily capturing spectral and spatial
information and reduces the amount of interaction time for the analyst.
1163 Improved Estimation of Environmental Parameters
through Locally Calibrated Multivariate Regression Analyses
Fabio Maselli
Abstract Download
Full Article
Linear uni- and multivariate regression analyses are commonly applied to relate
land surface parameters to the relevant spectral responses. In practice, this
is often the only means to extract operationally useful information from remotely
sensed data. The use of regression techniques over relatively wide areas is
however constrained by the spatial variability of the observed relationships,
which can originate from several causes. To overcome this problem, a modified
approach based on the local calibration of regression models is proposed. The
method, derivable from the fuzzy set theory, was originally introduced to enhance
the performance of conventional multivariate regressions applied to spatially
distributed data. The statistical bases of locally calibrated regressions are
first presented, together with an operational method to find the optimal model
configuration for each application. Two case studies are then described to
illustrate the performances of the locally calibrated multivariate regressions
compared to those of traditional procedures. The first case study, in particular,
exhaustively showed the potential and limitations of the new procedures to
extract climate parameters from mean monthly NOAA-AVHRR NDVI data. The second
case study dealt with the estimation of forest composition by the use of Landsat
TM images. Both investigations indicated that locally calibrated procedures
can produce more accurate predictive models than conventional regressions.
Additionally, these procedures can provide spatial estimates of accuracy statistics
which are useful for a better interpretation of the results and for subsequent
data integration.
1173 A Subpixel Classifier for Urban Land-Cover
Mapping Based on a Maximum-Likelihood Approach and Expert System Rules
Ming-Chih Hung and Merrill K. Ridd
Abstract Download
Full Article
A supervised classifier was developed to estimate ground component percentages
from Landsat TM (Thematic Mapper) images of urban areas. Six ground components
were selected according to the V-I-S (Vegetation-Impervious surface-Soil) model.
With aid from ERDAS/IMAGINE;rm software, implementation of this classifier
involved the Bayes algorithm, to calculate initial percentages, and expert
system rules, to iteratively adjust percentages according to a linear mixture
model. The end product is a six-channel image in which each channel indicates
percentages of a pre-defined ground component at the subpixel level. The resultant
image displays information beyond typical per-pixel classification results.
Pixels are still represented by six numbers, indicating the percentages of
six pre-defined ground components. Because the result is numerical and not
categorical, a more detailed accuracy assessment, other than an error matrix,
is needed. Therefore, a regression analysis was performed to compare the estimated
percentages to the surveyed percentages, which were derived from aerial photointerpretation.
Correlation coefficients were reported as indices of accuracy for each ground
component. This new technique was applied to a 1990 TM image covering portions
of the Salt Lake City area, Utah. Overall, the calculated indices of accuracy
show a significant relationship between the estimated and surveyed percentages.
Of the six correlation coefficients, two have strong relationships, three have
moderate relationships, and one has a weak relationship.
1181 Population Estimation Models Based on Individual
TM Pixels
Jack T. Harvey
Abstract Download
Full Article
There is a fundamental spatial mismatch in the data available for modeling
human population from satellite imagery. Spectral reflectances are available
for each pixel of an image, but ground reference population data are available
only for larger zones. The general response has been to build models for the
average population density of the zones, utilizing spatially aggregated spectral
data. This approach has limitations, both for the modeling process and for
the utilization of the resulting spatially aggregated population estimates.A
pixel-based alternative is described. Pixels of a Landsat TM image were classified
as residential or non-residential using standard techniques. Initial reference
populations were assigned by uniformly distributing the population of each
zone across its residential pixels. An expectation-maximization (EM) algorithm
was used to iteratively regress pixel population on spectral indicators and
re-estimate pixel populations. Predictive validity was tested by applying the
fitted regression equation to a second image. The pixel-based model produced
population estimates of comparable accuracy to those resulting from a much
more complex zone-based modeling procedure. The pixel-based procedure was also
more robust and more amenable to refinement, particularly at the extremes of
population density. The relative error in the estimated total urban population
of both primary and secondary study areas was less than 1 percent. Median relative
error in the population of individual zones was 16 percent in the primary study
area (14 percent for urban zones) and 21 percent in the secondary study area
(17 percent for urban zones).
1193 Estimating Residual Wheat Dry Matter from
Remote Sensing Measurements
Nereu Augusto Streck, Donald Rundquist, and Joel Connot
Abstract Download
Full Article
Plant material left in the field after harvest is an important product of agricultural
ecosystems. A laboratory experiment was carried out with the objectives of
(1) testing optical remote sensing as a means of measuring the amount of wheat
residue dry matter (DM), and (2) modeling the relationship between the amount
of wheat residue and reflectance indices. Six backgrounds (three types of soil
with dry and wet surfaces) and increasing amounts of wheat residue DM (0, 20,
50, 100, 150, 200, 250, 300, 400, 800, and 1500 g/m;s2) were used. A nonlinear
equation was used to estimate DM based on two reflectance indices, the Cellulose
Absorption Index (CAI) and the Normalized Difference Index (NDI). When CAI
was used as input, the model performed well with an average error of about
50 g/m2 and the residue was discriminated from soil up to a dose
of DM of about 600 g/m2. When NDI was used, the model was only able
to discriminate residue from soil when DM was less than about 300 g/m2,
with an average error of estimation greater than 100 g/m2. Limitations
of this study include the fact that residue moisture and weathering were not
considered.
1203 3D Model-Based Tree Measurement from High-Resolution
Aerial Imagery
P. Gong, Y. Sheng, and G.S. Biging
Abstract Download
Full Article
This paper introduces a 3D model-based tree interpreter, a semi-automatic method
for tree measurement from high-resolution aerial images. It emphasizes the
extraction of the 3D geometric information such as tree location, tree height,
crown depth (or crown height), crown radius, and surface curvature. First,
trees are modeled as 3D hemi-ellipsoids with the following parameters: tree-top
coordinates, trunk base height, crown depth, crown radius, and crown surface
curvature. This model-based approach turns a tree interpretation task into
a problem of optimal tree model determination. Multi-angular images are used
to determine the optimal tree model for each tree. Tree tops in each image
of a stereo pair are identified interactively with the epipolar constraint,
and the 3D geometry of trees can be determined automatically. With such a semi-automatic
scheme, efficiency and reliability of 3D tree measurements are achieved by
taking advantages of both the operator's interpretation skills and the machine's
computation. This paper mainly deals with conifers. The method was tested with
a closed conifer stand on 1:2,400-scale photographs. An overall accuracy of
94 percent and 90 percent was obtained for tree height and crown radius measurements,
respectively.
| Top | Home |