Peer-Reviewed Article Abstracts
483-490 A
Model to Support the Integration of Image Understanding Techniques within a
GIS
Mark Gahegan and Julien Flack
Abstract
Download
Full Article
Traditionally, geographic information systems (GIS) have used data generated
from remote sensing, but only after the data have been preprocessed in some
way to provide a suitable classification. This leads to several weaknesses;
the resulting framework is inflexible and cannot support multiple interpretations
of the same area. A new model for a GIS is introduced that includes a set
of image understanding methods, and expert knowledge governing the application
of these methods. The methods employed are chosen automatically to best emphasize
the types of features (rivers, fields, forests, etc) that the user is currently
investigating, and the type of imagery available. An implementation of the
model is described along with details of the knowledge structures and image
processing methods used. Examples are included to show the resulting adaptive
nature of image interpretation and subsequent inclusion of feature descriptions
into the GIS.
491-499 Relating the Land-Cover Composition of
Mixed Pixels to Artificial Neural Network Classifica-tion Output
Giles M. Foody
Abstract
Download
Full Article
Artificial neural networks are attractive for use in the classification of
land cover from remotely sensed data. In common with other classification
approaches, artificial neural networks are used typically to derive a 'hard'
classification, with each case (eg pixel) allocated to a single class. However,
this may not always be appropriate, especially if mixed pixels are abundant
in the data set. This paper investigates the potential to derive information
on the land- cover composition of mixed pixels from an artificial neural
network classification. The approach was based on relating the activation
level of artificial neural network output units, which indicate the strength
of class membership, to land-cover composition. Two case studies are discussed
which illustrate that the activation level of the artifical neural network
outputs themselves were not strongly related to pixel compsition. However,
re-scaling the activation levels, to remove the bias towards very high and
low strengths of class membership imposed by the unit activation function,
produced measures that were strongly related to the land-cover composition
of mixed pixels.
501-511 An Operational GIS Expert System for Mapping
Forest Soils
Andrew K. Skidmore, Fiona Watford, Paisan Luckananurug, and P.J.
Ryan
Abstract
Download
Full Article
The successful integration of a Bayesian expert system with a commercially
available geographic information system (GIS) (Genamap) is described. The
package mapped forest soils into five soil landscape classes by utilizing
a digital terrain model and vegetation map, as well as knowlege provided
by a soil scientist. It is concluded that the map produced by the expert
system was as accurate as the map drawn by the soil scientist, within a 95%
confidence interval. An overall mapping accuracy of 69.8% was achieved for
the soil maps produced by the expert system, while the conventionally derived
map had an accuracy of 73.6%.
513-523 Integrated Analysis of Spatial Data from
Multiple Sources: Using Evidential Reasoning and Artificial Neural Network
Techniques for Geological Mapping
P. Gong
Abstract
Download
Full Article
For classification purposes, commonly used remote sensing algorithms such as
the maximum-likelihood classifier and the minimum-distance classifier can
only be used to deal with spatial data of interval and ratio scales. They
are not applicable to spatial data of nominal or ordinal scale as exemplified
by data digitized from a categorical map. Bayesian theory, mathematical theory
of evidence, and artifical neural networks, on the other hand, are capable
of handling data with any measurement scale. In this paper, we introduce
an evidential reasoning and a back-propagation feed-forward neural network
algorithm and evaluate their applications to classification problems. A multisource
data set including Landsat Thematic Mapper, aeromagnetic, radiometric, and
gravity data has been used in the classification of four rock types. The
evidential reasoning method resulted in three highest individual class accuracies
out of the four classes.
525-531 Automatic Extraction and Evaluation of
Geological Linear Features from Digital Remote Sensing Data Using a Hough Transform
Arnon Karnieli, Amnon Meisels, Leonid Fisher, and Yaacov Arkin
Abstract
Download
Full Article
The Hough transform is an established tool for discovering linear features
in images. The present investigation presents a new and specific algorithm
for detecting geological lineaments in satellite images and scanned aerial
photographs which incorporates the Hough transform, a new kind of a 'directional
detector,' and a special counting mechanism for detecting peaks in the Hough
plane. Three test sites representing different geological environments and
remote sensing altitudes were selected. The first site represents sedimentary
conditions of chalk beds on cherry picker photography; the second represents
plutonic conditions of granite rocks on an aerial photograph; and the third
represents tectonic fractures of carbonates, chalks, and cherts on digital
satellite data. In all cases, automatic extraction and mapping of lineaments
conformed well to interpretation of lineaments by human performance.
533-538 Vision-Based Image Processing of Digitized
Cadastral Maps
Liang-Hwei Lee and Tsu-Tse Su
Abstract
Download
Full Article
This paper presents an automatic method for processing digitized images of
cadastral maps. The method includes two major algorithms: a segmentation
and a Raster-to-Vector conversion. Those algorithms use a simple data-list
structure for recording data created during single-pass, row-majority scanning
and line tracing. The segmentation algorithm obtains the positions and sizes
of symbols and characters, in addition to completing map segmentation and
proving useful for pattern recognition. The Raster-to-Vector conversion algorithm
obtains topological information necessary to relate cadastral map spatial
data to line start points, midpoints, intersection points, and termination
points. It consists of four integrated sub- algorithms that remove noise,
unify run-length coordinates, and perform synchronous line approximations
and logical linkage of line breaks. Straight, angled, and curved lines can
then be completely reconstructed for display. Also presented are six indices
that verify algorithm and experimental results.