Peer-Reviewed Articles
893 Optimizing Image Resolution to Maximize the Accuracy of Hard Classification
P.K. Bøcher and K.R. McCloy
Abstract Download
Full Article
There are three strategies by which the accuracy of classification
can be improved after the imagery that will be used
for the classification has been chosen. These are to improve
the definition of the class decision surfaces, to maximize the
between class distances, and to reduce the within class
variances. This paper reports on work done to investigate
the relationship between classification accuracy and within
class variances, where generally accepted measures of
accuracy derived from the Confusion Matrix are used as the
indicators of classification accuracy. This paper shows that
the within class variances are a function of image resolution,
and it provides a mechanism based on the Average
Local Variance (ALV) function to find the resolution that will
yield the highest relative within field classification accuracy
by minimizing the within class variances.
905 The Importance of Scale in Object-based Mapping of Vegetation Parameters with Hyperspectral Imagery
Elisabeth A. Addink, Steven M. de Jong, and Edzer J. Pebesma
Abstract Download
Full Article
In recent years, object-oriented image analysis has been
widely adopted by the remote sensing community. Much
attention has been given to its application, while the
fundamental issue of scale, here characterized by spatial
object-definition, seems largely neglected.
In the case of vegetation parameters like aboveground biomass and leaf area index (LAI), fundamental objects are individual trees or shrubs, each of which has a specific value. Their spatial extent, however, does not match pixels in size and shape, nor does it fit the requirements of regional studies. Estimation of vegetation parameters consequently demands larger observation units, like vegetation patches, which are better represented by variably shaped objects than by square pixels.
This study aims to investigate optimal object definition for biomass and LAI. We have data from 243 field plots in our test site in southern France. They cover a vegetation range from landes to garrigue to maquis, which is considered to be the climax vegetation in the area. A HyMap image covers the area.
The image is subjected to a Minimum Noise Fraction (MNF) transformation, after which it is segmented with ten different heterogeneities. The result is ten object sets, each having a different mean object size. These object sets are combined with the original image with the mean band values serving as object attributes.
Field observations are linked to the corresponding objects for each object set. Using Ridge regression, relations between field observations and spectral values are identified. The prediction error is determined for each object set by cross validation. The overall lowest prediction error indicates the optimal heterogeneity for segmentation.
Results show that the scale of prediction affects prediction accuracy, that increasing the object size yields an optimum in prediction accuracy, and that aboveground biomass and LAI can be associated with different optimal object sizes. Furthermore, it is shown that the accuracy of parameter estimation is higher for object-oriented analysis than for per-pixel analysis.
913 Integrating Fine Scale Information in Super-resolution Land-cover Mapping
Alexandre Boucher and Phaedon C. Kyriakidis
Abstract Download
Full Article
Super-resolution or sub-pixel class mapping is the task of
providing fine spatial resolution maps of, for example, landcover
classes, from satellite sensor measurements obtained
at a coarser spatial resolution. Often, the only information
available consists of coarse class fraction data, typically
obtained through spectral unmixing. This paper shows how
to integrate, in addition to such coarse fractions, class labels
at a set of fine pixels obtained independent of the satellite
sensor measurements. The integration of such fine spatial
resolution information is achieved within the Indicator
Kriging formalism in either a prediction or simulation mode.
The spatial dissimilarity or texture of class labels at the fine
(target) resolution is quantified in a non-parametric way
from an analog scene using a set of experimental indicator
semivariogram maps. The output of the proposed procedure
consists of maps of probabilities of class occurrence, or of a
series of simulated class maps characterizing the inherent
spatial uncertainty in the super-resolution mapping process.
923 Variability in Soft Classification Prediction and Its Implications for Sub-pixel Scale Change Detection and Super Resolution Mapping
Giles M. Foody and H.T.X. Doan
Abstract Download
Full Article
The impact of intra-class spectral variability on the estimation
of sub-pixel land-cover class composition with a linear
mixture model is explored. It is shown that the nature of
intra-class variation present has a marked impact on the
accuracy of sub-pixel class composition estimation, as it
violates the assumption that a class can be represented by a
single spectral endmember. It is suggested that a distribution
of possible class compositions can be derived from pixels
instead of a single class composition prediction. This
distribution provides a richer indication of possible subpixel
class compositions and highlights a limitation for
super-resolution mapping. Moreover, the class composition
distribution information may be used to derive different
scenarios of changes when used in a post-classification
comparison type approach to change detection. This latter
issue is illustrated with an example of forest cover change in
Brazil from Landsat TM data.
935 Assessing Alternatives for Modeling the Spatial Distribution of Multiple Land-cover Classes at Sub-pixel Scales
Yasuyo Makido, Ashton Shortridge, and Joseph P. Messina
Abstract Download
Full Article
We introduce and evaluate three methods for modeling the
spatial distribution of multiple land-cover classes at subpixel
scales: (a) sequential categorical swapping, (b) simultaneous
categorical swapping, and (c) simulated annealing.
Method 1, a modification of a binary pixel-swapping
algorithm, allocates each class in turn to maximize internal
spatial autocorrelation. Method 2 simultaneously examines
all pairs of cell-class combinations within a pixel to determine
the most appropriate pairs of sub-pixels to swap.
Method 3 employs simulated annealing to swap cells. While
convergence is relatively slow, Method 3 offers increased
flexibility. Each method is applied to a classified Landsat-7
ETM dataset that has been resampled to a spatial resolution
of 210 m, and evaluated for accuracy performance and
computational efficiency.
945 Scaling Field Data to Calibrate and Validate Moderate Spatial Resolution Remote Sensing Models
A. Baccini, M.A. Friedl, C.E. Woodcock, and Z. Zhu
Abstract Download
Full Article
Validation and calibration are essential components of
nearly all remote sensing-based studies. In both cases,
ground measurements are collected and then related to the
remote sensing observations or model results. In many
situations, and particularly in studies that use moderate
resolution remote sensing, a mismatch exists between the
sensor’s field of view and the scale at which in situ measurements
are collected. The use of in situ measurements for
model calibration and validation, therefore, requires a
robust and defensible method to spatially aggregate ground
measurements to the scale at which the remotely sensed
data are acquired. This paper examines this challenge and
specifically considers two different approaches for aggregating
field measurements to match the spatial resolution of
moderate spatial resolution remote sensing data: (a) landscape
stratification; and (b) averaging of fine spatial resolution
maps. The results show that an empirically estimated
stratification based on a regression tree method provides a
statistically defensible and operational basis for performing
this type of procedure.
Color Figures:
955 Land-use/Land-cover Classification with Multispectral and Hyperspectral EO-1 Data
Bing Xu and Peng Gong
Abstract Download
Full Article
We compared the capability of the Earth Observing-1 (EO-1)
Hyperion hyperspectral (HS) data with that of the EO-1
Advanced Land Imager (ALI) multispectral (MS) data for
discriminating different land-use and land-cover classes in
Fremont, California. We designed a classification scheme of
two levels with level I including general classes and level II
including more specific classes. Classification shows that the
HS data does not produce better results than the MS data when
we directly applied a Mahalanobis distance (MD) classifier.
We tested a number of feature reduction and extraction algorithms for the HS image. These algorithms include principal component analysis (PCA), segmented PCA (SEGPCA), linear discriminant analysis (LDA), segmented LDA (SEGLDA), penalized discriminant analysis (PDA) and segmented PDA (SEGPDA). Feature reductions were all followed by an MD classifier for image classification. With SEGPDA, SEGLDA, PDA, and LDA, similar accuracies were achieved while a segmentation- based approach we proposed (SEGPDA or SEGLDA) greatly improved computation efficiency. They all outperformed SEGPCA and PCA by 4 to 5 percent (level II) and 1 to 3 percent (level I) in classification accuracy.
For level II classification, overall accuracies obtained by using the features extracted from the HS image were 2 to 3 percent greater than those obtained with the MS image. For various vegetation class and impervious land use categories, the HS data consistently produced better results than the MS data. For level I classification, the HS image generated a thematic map that is 0.01 greater in kappa coefficient comparing to the MS image. When we collapsed the level II classification map to a level I map, 5 percent (HS) to 7 percent (MS) improvements were achieved.