Hyperspectral images of
natural scenes 2004
Scenes
The eight hyperspectral
images available
here were acquired from the same sites as those illustrated in Figure 1
of Foster,
Nascimento, and Amano, 2004,
and are a representative sample of the 25 images used
in that study. They consisted
of a
mixture of rural scenes from the Minho
region of
Portugal, containing, rocks,
trees, leaves, grass, and earth and urban scenes from the cities of
Porto and Braga, Portugal. Images were obtained during the
summers of
2002 and 2003, almost always under a clear sky. Particular care was
taken to avoid scenes containing movement. Scenes were
illuminated by direct sunlight in clear or
almost clear sky. Colour pictures of the eight representative
scenes
are shown here, and full-size
versions (each about 4MB) can be downloaded by clicking on each.
The estimated
reflectance spectra
(effective spectral reflectances) at each pixel in each of these
example scenes
can be downloaded as a WinZip file by clicking on the link at
the bottom of each picture. Each zipped file
contains contains for each
scene
a Matlab
scene-reflectances file (size e.g. 1018 x 1339 x 33), a Matlab
radiance-by-reflectances file for converting reflectances to radiances
(33 x 2), a Matlab reflectances file for the Munsell surface in the
scene (33 x 2), a bright BMP image of the scene, and a Microsoft RTF
information file. The version of Matlab used to save
these files was 7.2.0.232 (R2006a). Details of the
hyperspectral system, image acquisition, and
processing are given further down the page.
A DVD containing the
eight
representative scenes is also available on request to d.h.foster@manchester.ac.uk
or
smcn@fisica.uminho.pt.
If you use these images,
please cite this source publication: Foster, D.H.,
Nascimento, S.M.C., & Amano, K. (2004). Information limits on
neural identification of colored surfaces in natural scenes. Visual
Neurosci., 21, 331-336.
.
Details of hyperspectral
system, image
acquisition, and processing
The present system used a
low-noise
Peltier-cooled digital camera providing an x-y
spatial resolution of 1344 x
1024 pixels (Hamamatsu, model C4742-95-12ER, Hamamatsu Photonics K. K.,
Japan) with a fast tunable liquid-crystal filter (VariSpec, model
VS-VIS2-10-HC-35-SQ, Cambridge Research & Instrumentation,
Inc.,
MA) mounted in front of the lens, together with an infrared blocking
filter. Focal length was typically set to 75 mm and aperture to f/16 or
f/22 to achieve a large depth of focus. The line-spread function of the
system was close to Gaussian with standard deviation approx. 1.3 pixels
at 550 nm. The intensity response at each pixel, recorded with 12-bit
precision, was linear over the entire dynamic range. The
peak-transmission wavelength was varied in 10-nm steps over
400–720 nm
and the bandwidth (FWHM) was 10 nm at 550 nm, decreasing to 7 nm at 400
nm and increasing to 16 nm at 720 nm.
Before image
acquisition, the exposure
at each wavelength was determined by an automatic routine so that
maximum pixel output was within 86–90% of saturation.
Immediately
after
acquisition, the reflected spectrum from a small flat grey (Munsell N5
or N7) reference surface in the scene was recorded with a
telespectroradiometer (SpectraColorimeter, PR-650, Photo Research Inc.,
Chatsworth, CA), the calibration of which was traceable to the National
Physical Laboratory.
The raw images acquired
by the
hyperspectral camera were corrected only for dark noise, spatial
nonuniformities (mainly off-axis vignetting), stray light, and any
wavelength-dependent variations in magnification or
translation.
For all scenes, the signal at 400 nm was relatively noisy, but data for
this wavelength were not excluded. For Scene 5, however, the
signal at 720 nm was excessively noisy, and data for this
wavelength were excluded, so the
size of the scene-reflectances file is 1020 × 1339
×
32 (in any event, the signals at 400 and 720 nm make relatively small
contributions
to the in-vivo response of human cones to natural reflected
spectra).
Edited versions
(labelled "_lax") have
also been included for Scenes 1, 2, 3, and 5. These edited
versions
have been corrected for some nonlinear chromatic artefacts arising from
movement of foliage within the scene during image
acquisition.
Effective spectral
reflectances at each
pixel were estimated by
normalizing the corrected signal against that obtained from the grey
reference surface in the scene. An informal
description of effective spectral
reflectances and the effective global illuminant are given below. Further technical
details are
given in Foster
et
al. (2004)
and in Foster
et
al. (2006),
the latter including additional calibration data (p.
2360) and an
analysis of
effective spectral reflectances with direct and indirect illumination (Appendix
A, pp.
2370-2371).
These data are for personal use
only. Acknowledgement of the relevant source publications should be given in any
published work arising from these data: Nascimento,
Ferreira, and
Foster, 2002, Foster
et
al. (2004), or
Foster
et
al. (2006).
Notes on effective
spectral reflectances
and effective global illuminant
The effective spectral
reflectance at
each pixel was obtained by dividing the spectral radiance
recorded by the hyperspectral camera at that position and wavelength by
the spectral radiance recorded from a calibrated neutral reference
surface (a Munsell grey) in the scene and then multiplying by the known
spectral reflectance of the reference surface. The reference surface
was usually flat and placed vertically facing the camera. Because
surfaces oriented at an angle to the camera may reflect more light than
vertical surfaces, their effective spectral reflectances may exceed
unity.
There are three ways to
deal with this.
1. If there is a Munsell
grey sphere in
the scene
with a visible highlight, then multiply (wavelength-by-wavelength) the
effective spectral reflectance at each
pixel by the effective spectral reflectance
at the flat reference and then divide by the effective spectral reflectance
at the highlight. If there is a true specular highlight in the scene,
e.g. from water (Scene 4, top right-hand corner), then this might
instead be used as the divisor
or it can be removed altogether from
the calculation.
2. Divide all the
effective spectral
reflectances in the scene by a constant equal to the maximum effective
spectral reflectance evaluated over all pixels and wavelengths.
3. Ignore it. The
effective spectral
reflectances are normally used to calculate a reflected spectrum.
Multiplying the effective spectral reflectance by the data in the
radiance-by-reflectances file recovers the original spectral radiance
recorded by the camera. Multiplying instead by a new different
illuminant spectrum will simulate the effect of replacing the original
illuminant by this new illuminant. Thus there is a trade-off between
decreasing all the spectral reflectances in the scene by a constant
factor, as in method 2, and increasing the illuminant spectrum by the
same factor.
The spectrum of the
effective
global
illuminant was obtained by taking the actual reflected spectrum from a calibrated
neutral reference surface (a
Munsell grey) in the scene recorded with a
telespectroradiometer and
dividing by the known spectral reflectance of that surface. The
spectral
radiance recorded by the hyperspectral camera at each position may then
interpreted as a product of this effective global illuminant and the
effective spectral reflectance at that position.
Further technical
details are
given in Foster
et
al. (2006), particularly Appendix
A, pp.
2370-2371.
Tutorial on transforming
hyperspectral
images
Click here
for a tutorial on transforming hyperspectral image reflectances into
reflected radiances and RGB colour images.
The tutorial includes sample code
(MATLAB, The MathWorks Inc) and technical notes on the interpretation
of reflectance spectra from natural scenes. Also available for download
is a zipped package containing a small test hyperspectral image,
daylight illuminant spectra, and a conversion routine for producing
sRGB images.
(c)
D. H. Foster, S. M. C. Nascimento, and K. Amano, March 2011
[Return
to home page]
