Full text: Fusion of sensor data, knowledge sources and algorithms for extraction and classification of topographic objects

International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
Bruno Aiazzi 1 , Luciano Alparone 2 , Stefano Baronti 1 , Ivan Pippi 1 
^‘Nello Carrara” Institute on Electromagnetic Waves IROE-CNR, Via Panciatichi, 64, 50127 Firenze, Italy, baronti@iroe.fi.cnr.it 
“Department of Electronic Engineering, University of Firenze, Via S. Marta, 3, 50139 Firenze, Italy, alparone@lci.die.unifi.it 
KEYWORDS: Data Fusion, Multispectral Images, Landsat TM, MOMS-2P, Multiresolution Analysis, Generalized Laplacian Pyramid 
Goal of this work is to present a general and formal solution to the problem of synergetic integration of multi-sensor image data, 
which may be collected with practically whatsoever spectral and ground resolution, although scale ratios larger than, say, 10, may 
be questionable in certain applicative contexts. The proposed data fusion methodology is applied to a specific example concerning 
observations from two different multi-spectral satellite sensors: Landsat TM (30 m resolution) and MOMS-2P (18 m ). The fusion 
procedure relies on the generalized Laplacian pyramid (GLP), which is a non-octave band-pass analysis structure unconstrained from 
the ground scales of the imaged data. The advantage is that the base-band extracted from the finer image exactly matches both in 
displaying scale and in resolution, i.e. content of spatial frequencies, the coarser image. For any rational scale ratio, a unique low-pass 
filter is needed. The filter design, however, is easy and noncritical for performances. The GLP scheme is superior in enhancing spatial 
features while retaining multi-spectral signatures, according to both objective and subjective quality criteria. The experiment reported 
highlights the assets of the pyramid approach: 5:3 fusion is designed for spectral enhancement of MOMS-2P (three 18 m bands) 
through Landsat TM (six 30 m bands). Since the three available multi-spectral bands of MOMS-2P overlap with Bands 2, 3, and 4 of 
TM, the test consists of synthesizing the three missing MOMS-2P bands -one in the blue and two in the infrared wave-lengths- from 
the TM data, in order to obtain 18m R-G-B true colour and three 18 m bands in near-medium infrared. 
The ever increasing availability of space-borne sensors imag 
ing in a variety of ground scales and spectral bands makes multi 
sensor fusion of multi-spectral data a discipline to which more 
and more general formal solutions to a number of applications 
are demanded. Since physical constraints impose a trade-off 
between spatial and spectral resolution, spatial enhancement of 
poor-resolution multi-spectral data is desirable, as well as spec 
tral enhancement of data collected with adequate ground resolu 
tion but poor spectral selection. 
After data registration and resampling to obtain a homoge 
neous spatial scale, image data fusion may be approached at three 
levels of increasing hierarchy (Pohl, 1998): 
• Pixel-level combination: 
- deterministic: Intensity-Hue-Saturation (IHS) transform; 
- statistical: Principal Components Substitution (PCS). 
• Feature-level combination: 
- High-Pass Filtering (HPF) method (Chavez, 1991); 
- ARSIS method based on wavelet transform (Ranchin, 1996); 
- Generalized Laplacian pyramid (GLP) (Aiazzi, 1998). 
• Decision-level combination of pixel values and/or features 
based on previously extracted information. 
In applications of classification, however, merging algorithms 
are often requested to maintain the spectral characteristics of the 
original data (Wald, 1997). In this light, the HPF method is more 
efficient than pixel-level algorithms (IHS and PCS) in preserv 
ing spectral features of spatially enhanced bands (Chavez, 1991). 
HPF consists of an injection of high-frequency components into 
resampled versions of the multi-spectral data. Such components 
are extracted by taking the difference between a higher resolution 
observation and its low-pass version originally obtained through 
a box filter, i.e. a square digital filter with equal coefficients. 
The idea of HPF was developed in a formal multi-resolution 
framework by employing the wavelet transform (WT) as 
analysis-synthesis tool (Garguet-Duport, 1996; Li, 1995; Nunez, 
1999; Yocky, 1996; Zivco, 1998). Perhaps, the rationale ex 
pressed in HPF was best embodied by the ARSIS method 
(Ranchin, 1996), which gives excellent results when the scale 
ratio of the data is a power of 2 (Wald, 1997), but experiences 
some troubles otherwise (Blanc, 1998), because of non-flexible 
structure of the WT, which requires a filter-bank, i.e. a couple of 
filters, one low-pass and another high-pass, having complemen 
tary symmetry of responses, whose design is non-trivial. 
A different multi-resolution approach was pursued in the con 
text of Laplacian pyramids (LP) (Burt, 1983, 1984; Baronti, 
1994; Aiazzi, 1997b) by other researchers, including the authors 
(Toet, 1989; Wilson, 1997; Aiazzi, 1997a, 1997c, 1998). Fol 
lowing the LP structure, an image is decomposed into nearly dis 
jointed band-pass channels in the spatial frequency domain, with 
out losing the spatial connectivity of its edges. The LP can be 
easily extended to deal with scales whose ratios are rational num 
bers, not only powers of two, as for the WT. Specifically, for a 
rational scale ratio (> 1) a unique low-pass filter is needed. Fur 
thermore, filter design is easy and non-critical for performances 
(Aiazzi, 1997a). Thus, data collected with ground resolutions of, 
say, 3 m, 5 m and 7 m can be easily compared and merged. 
When spatial enhancement was concerned, the GLP scheme 
has been shown to be superior in retaining multi-spectral sig 
natures (Aiazzi, 1998) also according to objective quality crite 
ria (Wald, 1997). However, fusion of two sets of multi-spectral 
data having different, and complementary, number of bands and 
ground resolution, may be also approached as a problem of spec 
tral enhancement of the higher-resolution data. In this paper, 
we wish to describe an example highlighting how the method is 
simple to be designed and generalized to images having a vari 
ety of ground resolutions, as future applications concerning new- 
generation sensors will demand in a close future.

Note to user

Dear user,

In response to current developments in the web technology used by the Goobi viewer, the software no longer supports your browser.

Please use one of the following browsers to display this page correctly.

Thank you.