A FRAMEWORK FOR AUTOMATIC LOW-RESOLUTION SATELLITE IMAGE
INTERPRETATION BASED ON SPECTRAL, CONTEXTUAL
AND MULTITEMPORAL KNOWLEDGE
G. L. A. Mota *, K. Pakzad ", S. Muller *, M. S. P. Meirelles “°, R. Q. Feitosa **, H. L. da C. Coutinho*
“ Dept. of Electrical Engineering, Pontifical Catholic University, Rio de Janeiro, Brazil — (guimota, raul)@ele.puc-rio.br
° Institute of Photogrammetry and Geolnformation, University of Hannover, Germany — pakzad@ipi.uni-hannover.de
* Institut für Theoretische Nachrichtentechnik und Informationsverarbeitung, University of Hannover, Germany —
mueller@tnt.uni-hannover.de
? University of the State of Rio de Janeiro, Rio de Janeiro, Brazil — maggie@eng.uerj.br
* EMBRAPA Soils, Rio de Janeiro, Brazil — heitor@cnps.embrapa.br
Commission IV , WG IV/7
KEY WORDS: Environment, Agriculture, Automation, Knowledge base, Fuzzy Logic, Multitemporal
ABSTRACT:
This work proposes a framework for increasing the degree of automation of low-resolution satellite images interpretation procedures.
Basically, the method starts with an image segmentation which, according to a criterion of spectral response homogeneity, outlines
the regions to be classified. The classification procedure is aided by expert knowledge. This procedure makes use of 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 region and the plausible class transitions in that particular time interval. This
strategy takes simultaneously into account spectral, contextual and multitemporal evaluations of the region which, are combined into
a single membership value. Each membership value corresponds to a class of interest, and the highest indicates its classification. As
a consequence, the proposed model requires as input: satellite images of the region of interest acquired in different dates; the
accurate classification of the former image and the previous mentioned categories of expert knowledge. The prominent points of the
proposed methodology are: its flexible structure, which allows for straightforward application of this model to low-resolution image
interpretation cases; and the automatic learning/calibration of the spectral levels to the current time image. Experiments were
performed in order to evaluate the potential of the proposed framework. The images employed in the experiments are situated in the
Taquari Watershed, more exactly, in the County of Alcinópolis that belongs to the State of Mato Grosso do Sul, Brazil. The
experimental results indicate that the use of knowledge can contribute to the increment of the degree of automation of the
interpretation process
1. INTRODUCTION
Considering forests, ecosystems, productive areas and urban
areas, remote sensing constitutes a significant tool for the
monitoring of large regions of the surface of the Earth.
Namely, such technology can play an important role for the
management of the natural resources.
However, most software packages for remote sense data
interpretation are based on visual analysis. Although such
packages offer digital image processing and pattern
recognition tools, they lack of a framework to automate the
image interpretation procedure which could increase the
productivity of the photo-interpreters. Additionally, despite
software development allowing the automation of some
tasks, high level structuring is limited. Consequently, in this
field, the development of technology is essential.
The other motivation for this research derives from
economical restrictions. The cost of aerial or high resolution
satellite images — e.g. IKONOS -, considering the huge
dimensions of threatened ecosystems currently under
monitoring programs in countries like Brazil (tropical
rainforests, savanna and wetlands) monitoring, is too high.
This constitutes an obstacle to its utilization on a routine
basis.
Visual (manual) interpretation and analysis dates back to the
early beginnings of remote sensing for air photo
interpretation. On the other hand, digital processing and
analysis is more recent, it was brought by the advent of
digital recording of remote sensing data and the development
of computers. Both manual and digital techniques for
interpretation of remote sensing data have their respective
advantages and disadvantages. Consequently, by and large,
the interpretation of remote sensing data encompasses a
sequence of manual and automatic steps. Considering both
knowledge background and additional data, in the manual
steps, a photo-interpreter calibrates/trains the automatic
algorithms and also solves the inconsistencies of their
outcomes.
In low spatial resolution satellite images, distinct land cover
classes may produce similar spectral responses turning more
difficult their discrimination. Nonetheless, it does not
represent a problem to experienced human photo-interpreters
who take advantage of supplementary information (e.g. size,
shape and texture) as well as their knowledge background in
order to solve contradictory interpretations. The present
proposal considers that by explicitly modeling the experience
of a photo-interpreter about a specific site into a knowledge-
basis, his reasoning can be computationally reproduced.
Inter
—
This
so-Cé
insti
inter
imag
The
aims
algoi
hom
selec
spec
prev
and
takir
unde
the [
of tl
train
auto:
The
proc
class
know
of sy
cont
disci
mult
class
trans
This
disci
inter
secti
The
been
Syste
Clen
main
The
COM]
scen
proc
play:
abou
In tl
and
some
