AGE
e.puc-rio.br
innover.de
rmany —
1 procedures.
city, outlines
ree types of
5; contextual,
| reasoning,
iterval. This
ymbined into
ification. As
it dates; the
points of the
lution image
iments were
tuated in the
Brazil. The
ation of the
| back to the
air photo
cessing and
e advent of
levelopment
hniques for
r respective
y and large,
ompasses a
dering both
the manual
> automatic
es of their
t land cover
urning more
= does not
-interpreters
n (e.g. size,
ckground in
he present
' experience
knowledge-
]uced.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004
This research is part of an international cooperation project,
so-called ECOWATCH, involving Brazilian and German
institutions, aiming at setting up systems for the automatic
interpretation of multitemporal low-resolution | satellite
images.
The visual (manual) interpretation process that this work
aims at automating starts with an image segmentation
algorithm which outlines contiguous segments with
homogeneous spectral response; then, the photo interpreter
selects some segments as a training set for the learning of
spectral signatures. Finally, considering simultaneously the
previous classification of one segment, its spectral pattern
and the context where it is situated, the photo-interpreter,
taking also into account his knowledge about the region
under analysis, classifies the segment. The main objective of
the present proposal is to improve the degree of automation
of the process as a whole; therefore, the selection of the
training set and knowledge-based classification must be
automated.
The proposed framework starts with an image segmentation
procedure, as explained above. A knowledge-based
classification engine is employed, considering three types of
knowledge: spectral, which relates the homogeneous classes
of spectral response to the correspondent classes of interest;
contextual, indicating the relevant contexts for the
discrimination of classes with similar spectral responses; and
multitemporal reasoning, considering both the former
classification of the segment and the plausible class
transitions in that particular time interval.
This paper is organized as follows: The next section
discusses the state of the art of knowledge based image
interpretation. Section 3 presents the proposed methodology,
section 4 the experiments and section 5 the conclusions.
2. KNOWLEDGE BASED INTERPRETATION
The automatic interpretation of remotely sensed images has
been intensively researched. By analyzing the existing
systems available in the literature (Matsuyama, 1990;
Clement, 1993; Niemann, 1990, 1997; Bueckner, 2001) their
main components can be identified as follows:
a) Digital images from particular sensors,
b) GIS data of the focused region,
c) image processing algorithms,
d) prior knowledge on the focused region, and
¢) a control logic to the interpretation process.
The control logic manages the interaction of the remaining
components. This component triggers, accordingly to the
scene semantic, the image processing algorithms. In this
process, the prior knowledge, usually delivered by an expert,
plays an important role, by supplying specific information
about the expected objects.
In the literature, many approaches for image interpretation
and sensor fusion have been presented; nevertheless, only
some try to formalize the expert knowledge. Some cases, like
SPAM (McKeown, 1985) and SIGMA (Matsuyama, 1990).
implement a hierarchical control and construct the objects
incrementally, considering multiple levels of detailing.
MESSIE (Clement, 1993) models the objects explicity
distinguishing four views: geometry, radiometry, spatia!
context and functionality. It employs frames and production
rules. ERNEST (Niemann, 1990, 1997) applies semantic
networks in order to represent the structure of the objects,
serving as a knowledge base specific to the problem.
The increasing amount of regions represented in Geographic
Information Systems (GIS) motivated the development of
AIDA (Liedtke, 1997). It is able, in a single semantic
network, to model the expert knowledge, GIS information
and data from multiple sensors. Among other advantages, the
incorporation of GIS information reduces the interpretation
uncertainty. AIDA was applied to the analysis of aerial
images, where the image components like buildings, houses,
rivers, factories, and forests may be observed.
The semantic network into AIDA is organized in several
layers. The highest layer provides the semantic of the objects
foreseen, whilst the lowest level corresponds to the image
primitives. As a whole, the several layers correspond to
distinct abstraction levels, providing a structural description
of the scene.
Employing a novel modality of scene description, GEOAIDA
(Bueckner, 2001) incorporates a holistic approach to the
main advantages of its predecessor. It considers an object as
a whole, in a global way, in other words, without subdividing
it into its subcomponents. In its implementation, holistic
operators may be easily incorporated to the semantic network
nodes. Rigorously, GEOAIDA provides a hybrid approach, in
cases where the holistic operators are unable of acting a
structural analysis proceeds. The main advantage obtained
when the holistic operators perform properly is the reduction
of the amount of time spent by the knowledge interpretation,
which is such a time expansive process.
Considering the previously mentioned systems, their main
divergence corresponds to the formalization of the expert
knowledge and information acquisition. Even though
historically formalisms of knowledge representation had
been developed in order to process natural language, they are
quite versatile.
Little is reported in the literature about the use of knowledge-
based approaches to the interpretation of low-resolution
satellite images. Zhang (1998) described a method which
aims at detecting changes in the land use. The approach
simultaneously employs SPOT and LANDSAT images to
update the urban maps of Shanghai, China, and discriminates
vegetation, water and urban areas. As a result, the urban
maps are updated highlighting the new constructions.
Kunz et al. (1997) employed ERNEST to update the maps in
a GIS database. The approach derives a semantic network
from the contents of the GIS database. Beside the spectral
response, the compactness, the mean curvature, the texture
standard deviation and homogeneity are evaluated, and
compared with the contents of the GIS database.
Discrepancies are corrected, being the GIS updated.
Largouet et al. (2001), implemented a land cover analysis
using a sequence of images of different satellites and a