] OWS are
d by SSM/.
empirically
sible, if the
d out using
algorithm is
all possible
ed study in
tor and Dr.
r consistent
| to GHRC-
"wal, V.K.,
/I and IRS-
gion, Proc.
e, 5-8 Dec,
Microwave
ation from
nd Neural
Honolulu,
eophysical
rom IRS-
e Sensing
ia, pp 207-
and Rana,
/ Scanning
actions on
508.
L.K., 2002,
rain over
Geophys.
M., 2002,
1 oceanic
ults, Jour.
B.S., and
4-MSMR
d TRMM-
Earth and
, Mathur,
surements
nd. Acad.
'.K., 2002
ater from
submitted
or Special
, (C4), pp
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002
CONTENT-BASED BROWSING OF
HIGH SPATIAL RESOLUTION SATELLITE IMAGERY
A. Senthil Kumar **, D. K. Jain”, S. Chowdhury? and K.L. Majumder”
* Data Processing Area, National Remote Sensing Agency, Hyderabad 500037 — senthilkumar_a@nrsa. gov.in
? Signal and Image Processing Group, Space Applications Centre, Ahmedabad 380015 — {dinesh,klm,santanu} @ipdpg.gov.in
Commission VII, WG VII/4
KEY WORDS: High spatial resolution imagery, Image browsing, Gabor wavelet, Object identification.
ABSTRACT:
Due to high information cluttering, browsing an urban scene data from high-spatial resolution satellites for a specific object of
interest becomes quite cumbersome. À content-based browsing (CBB) method is proposed to circumvent this problem in this paper.
The proposed method consists of two major steps. The first step makes a coarse search of plausible regions in a given image
containing the object signature. A histogram intersection method is employed here for comparing histograms of the query object (in
memory) and of the search window in the given image. In the second step, feature vectors derived by using the Gabor wavelet
transform of the query object and search window data are compared. The performance of these feature vectors is also studied for
recognizing the object irrespective of rotation, scale, and gray level variations. The proposed method has been applied over the
IKONOS Panchromatic (spatial resolution — 1 m.) and over the /RS-1D Panchromatic (5 m.) satellite data, and found to be working
quite satisfactorily.
1. INTRODUCTION
In this paper, we address the problem of identifying objects
from satellite imagery of high spatial resolution (HSR). This
work assumes much relevance from viewpoint of recent
launches of several satellites with spatial resolutions ranging
from 5 meters to less than 1 meter. Due to high information
cluttering typical of the HSR data especially over urban areas, it
is found extremely difficult to locate an object of interest. It is
generally observed that it takes painfully long time for visual
browsing these imagery for the objects unless their geometric
locations are unknown a priori.
There have been many methods proposed in the related area of
automatic object recognition, but these are largely restricted to
night vision systems employing infrared cameras [1,2]. In the
infrared, targets such as%anks, trucks, ships etc. have strongly
different signature as compared to background, and are, thus,
less congested to identify. In contrast, the objects in the visible
HSR imagery are highly congested with background of similar
reflectance, thus making the visual identification quite difficult.
Recently, Manjunath and Ma [3] suggested the use of Gabor
filters for fast browsing of HSR air-borne imagery. The Gabor
filters are tolerant to distortions to a great extent between the
query object (in memory) and the test object to be recognized
from the given image. These distortions are typical of the HSR
satellite imagery due to high orbital maneuvering and
calibration differences of sensors from pre-flight
measurements.
In this paper, we describe a fast content based browsing (CBB)
scheme by combining a histogram-based search method
followed by the Gabor wavelet transform method. The
proposed approach is discussed in detail in Sec 2. In Sec.3, we
describe results of our work with an /RS-1D Panchromatic data
and with an IKONOS imagery with spatial resolution of 1 meter.
Finally, our conclusions are given in Sec. 4.
2. THE PROPOSED APPROACH
The scheme proposed here for the HSR satellite data is as
shown in Fig. 1. It basically consists of two modules, one
making a fast search for quickly providing regions in the image
which may possibly contain the object under search, and the
other analyzing these regions with Gabor wavelet features for
distortion invariant identification. The first search-level,
labeled in Fig.1 as SL-1, involves noise filtering followed by a
histogram matching based image segmentation step.
HSR image
Query
4 data
Image Partitioning
| Search windows
Background SL-1
Normalization
Y
Histogram
intersection
A
Selected Vv SL-2
wi ndo wa
Gabor feature
extraction
Y
Merging over- «
lapped patches
Object dra windows
Fig. 1. Proposed method for the HSR image browsing.
