static-content.springer.com10.1007... · Web viewdata cube norm = dc raw - df av wf av - df av * t int (white) t int (sample) dc raw = Data cube of raw data df av =Dark reference
Post on 22-Mar-2018
231 Views
Preview:
Transcript
Supplementary Online MaterialTable of content:
- Glossary: Hyperspectral Imaging
- SOM_Table 1: metrics, formula, description and references for the parameters used in
the text
- SOM_Fig. 1: Sediment core of Lake Jaczno with stratigraphic cluster zones and zoom-
in locations of SOM_Fig. 2
- SOM_Fig. 2: Zoomed images of pigment distributions and classification results from
two selected locations in the sediment core
- SOM_Fig. 3: correlation matrix for the multiproxy dataset (Fig 4) showing the
individual scatterplots, the Pearson correlation and significance levels
- References
Glossary: Hyperspectral imaging analysis
The following section contains a short glossary and description of the methods used for
hyperspectral image analysis. Note that there is a detailed manual for hyperspectral imaging
of lake sediments in Butz (2016). The extension package for the ENVI program can be
obtained from the corresponding author.
Normalization (Butz et al. 2015)
Normalization describes the transformation of camera raw data (radiometric counts) into
reflectance values using white- and dark references. Reflectance is the fraction of light
reflected from a sample with respect to the light source.
The normalization is calculated as:
data cubenorm=
dcraw−df av
wf av−df av∗t∫(white)
t∫ (sample )
dcraw = Data cube of raw data
dfav =Dark reference averaged into one frame
1
wfav = White reference averaged into one frame
tint = Integration time / exposure time
Median filter (Tukey 1977)
In a window of 5x5 spatial pixels the median value is calculated. Then the centre pixel is
replaced by the median value. Subsequently, the window is moved by one pixel and the
calculation starts again on the original dataset. This operation is performed in the spatial
dimension on all pixels in each spectral band. The result is a dataset where erroneously high or
low single pixels are smoothed. The first and the last two rows/columns of pixels located at
the edges of the dataset are kept in the original form.
Subsetting (Butz et al. 2015)
The image data is cut spatially to the extent of the sediment without the core liner or other
materials. Bad spectral bands (here: 396-499 nm) are removed.
Endmember detection (Butz et al. 2015; Kruse et al. 1999)
A spectral endmember is a single spectrum of a pure substance or a mixture of substances if
the pure form is not existent. Spectral endmembers are, therefore, the purest compounds in a
sample. All other spectra are combinations of the spectral endmembers.
We use the spectral hourglass wizard of the envi 5.03 software (Exelisvis ENVI, Boulder
Colorado) for endmember detection. This wizard performs a noise whitening of the
reflectance data by application of a principal component analysis (PCA). Then a second PCA
is applied to spectrally reduce the dataset. These two cascaded PCAs are called a Minimum
Noise Fraction (MNF). Then the dataset is spatially reduced by a pixel purity index (PPI).
Random vectors are projected on each combination of principal components. Pixels located at
extreme positions are counted until all extreme pixels are found. The remaining extreme pixels
are shown in an n-dimensional scatterplot and endmembers are selected from extreme
positions in the data cloud.
The figure shows endmember spectra from Lake Jaczno (Butz et al. 2015). Some spectra were
highlighted for better contrast. These endmembers are spectra of compound substances rather
than pure substances. Endmembers are derived in order to evaluate the range of spectra to be
2
found in a sample. Based on spectral features such as chlorophyll absorption (green bar) and
bacteriopheophytin absorption (blue bar) spectral indices are chosen and calculated.
Endmember detection (continued)
Spectral endmembers from Lake Jaczno (after Butz et al. 2015)
Spectral index calculation (Butz et al. 2015; Rein and Sirocko 2002)
The endmembers are investigated for absorption bands or prominent spectral features.
Depending on the features found, spectral indices are calculated. In the case of Lake Jaczno
(this paper), there are absorption features with minima at R673 and at R845 nm (Figure above
and SOM_Fig. 2). The spectral indices for the relative absorption band depth (RABD) were
calculated after Rein and Sirocko (2002). This method was adapted to the hyperspectral
camera and the absorption features of the spectral endmembers.
Spectral indices were calculated as:
RAB D673=(36∗R589+54∗R729
90)/R673
RAB D845=(34∗R790+34∗R899
68)/ R845
Ri = Reflectance at wavelength i in nanometers
RMean=1n∑i=0
n−1
Ri
3
Ri = Reflectance at spectral band index
n = Number of spectral bands
Minimum distance (Richards and Jia 1999)
This classification method uses the Euclidian distance between a sample spectrum and a range
of reference spectra (classes) to determine to which class the sample is closest. With no
thresholds set, every sample pixel is assigned a class.
The minimum Euclidian distance is calculated as:
Di ( x )=√( x−mi )T− ( x−mi )
D = Euclidian distance
I = i-th class
X = number of dimensions (image bands)
Mi = mean spectrum of class i
For Lake Jaczno, eight training classes were chosen from specific sediment regions (figure:
class means of training areas). Training areas were selected manually after thin section
analysis for non-varved areas and calcite, charcoal, clay and organic varve layers. RABD673
and RABD845 classes were created based on the highest RABD values for each class.
4
Class means of training areas. Bad bands (396-500 nm) were removed before classification.
Spectral angle mapper (Kruse et al. 1993)
The spectral angle mapper compares the calculated angle between the vector of a target
spectrum and the vector of a reference spectrum. If the angle is below a specified threshold
(this paper: 0.04 rad) then the sample classification is positive. If the angle is higher than the
threshold, the classification is negative.
The spectral angle mapper is calculated as:
α=cos−1 { ∑i=1
n
ai ∙ bi
√∑i=1
n
ai2 ∙√∑i=1
n
b i2 }
α = angle [rad]
a = target vector/spectrum
b = reference vector/spectrum
n = number of spectral bands (dimensions)
5
i = i-th band
The classes from the spectral library (see minimum distance classification, figure above) were
used to perform the classification.
6
SOM Table 1: metrics, formula, description and references for the parameters used in the
text.
Parameter Description References
Water content
[%]
w c=(mw−md
mw)∗100
wc = water content [%]
mw = mass of wet sediment [g]
md = mass of freeze-dried sediment [g]
(Menounos 1997)
Dry bulk density (DBD)
[g cm-3]
The DBD is the mass of the freeze-dried sediment
per volume.
Sedimentation rate (SR)
[cm yr-1]
The sedimentation rate is the varve thickness
(derived from varve counting).
Mass accumulation rate
(MAR)
[g c m−2 y r−1 ¿
MA R ( gcm−2 y r−1 )=DB D ( g c m−3 ) x S R (cm yr−1) (Zolitschka 1998)
Loss on ignition at 550°C
(LOI550) [%]
LOI550;FLUX [g cm-2 yr-1]
LO I 550 (% )=(md−md ( 550 ))/md ¿∗100
md = mass of freeze-dried sediment [g]
md(550) = mass after 4h @550°C
LO I 550; FLUX=(LO I550 (% )
100)x MA R( g c m−2 yr−1)
(Heiri et al. 2001)
(Zolitschka 1998)
Loss on ignition at 950°C
(LOI950) [%]
LOI950;FLUX [g cm-2 yr-1]
LO I 950 (% )=(md (950)−md ( 550 ))
md∗100
md = mass of freeze-dried sediment [g]
md(550) = mass after 4h @550°C
md(950) = mass after 2h @950°C
LO I 950 ;FLUX=(LO I 950 ( %)
100) x MA R ( g cm−2 y r−1 )
(Heiri et al. 2001)
(Zolitschka 1998)
Lithogenic flux (LF)
[g c m−2 y r−1 ¿ LF=(100−LO I 550 ( %)−LO I 950 (% ))
100∗MAR
Bacteriopheophytin a flux
[µg cm-2 yr-1]Bphe a( µgc m−2 yr−1)=Bphe a (µg g−1)∗MAR After Zolitschka
(1998)
7
Total chlorophylls flux
[µg cm-2 yr-1]Tchl( µg cm−2 y r−1 )=Tchl( µg g−1 )∗MAR After Zolitschka
(1998)
Supplementary Online Figures
8
SOM_Fig. 1: Lake Jaczno sediment core. The figure shows the stratigraphic cluster zones and a more detailed view on
the sediments. Coloured boxes show sections enlarged in SOM_Fig. 2.
9
SOM_Fig. 2: Section zooms from SOM_Fig. 1. A: Zoom-in to the top section of the sediment core showing the transition
from charcoal enriched layers to the non-charcoal state. The figure shows from the left to the right a true colour image,
the pigment distributions for Bphe a and total chlorophylls and the carbonate, the charcoal and the clay classifications.
Bphe a is completely absent while Tchls are weak. Carbonate layers are very thin in this section. The charcoal
classification shows a hard transition over one or two years. Clay classification shows ongoing erosion after the charcoal
was washed out from the catchment. The clay classification is spectrally more delicate than the pigment abundance.
Therefore, the clay classification is absent within the charcoal enriched layers because the charcoal signal is superseding
it. B: Zoom-in to the bottom section of the sediment core. The index setup is the same as in A, however, there were no
charcoal layers in the bottom section, and thus, the charcoal map was omitted. The figure shows the fine lamination of the
varves in the bottom section and the structure of the pigment distributions. A prominent clay layer can be observed at
~2130 mm sediment depth (greyish layer). The layer was well detected by the spectral angle mapper algorithm.
10
SOM_Fig. 3: Correlation matrix for the multiproxy dataset (Fig 4) showing the individual scatterplots and the Pearson
correlation and significance levels (*p<0.05,**p< 0.01,***p< 0.001).
11
References:
Butz C (2016) Hyperspectral imaging of lake sediments: Methods and applications in a
meromict lake of NE Poland. Institute of Geography & Oeschger Centre For Climate Change
Research. University of Bern, Bern, p 225
Butz C, Grosjean M, Fischer D, Wunderle S, Tylmann W, Rein B (2015) Hyperspectral
imaging spectroscopy: a promising method for the biogeochemical analysis of lake sediments.
J Appl Remote Sens 9:096031-096031
Heiri O, Lotter AF, Lemcke G (2001) Loss on ignition as a method for estimating organic and
carbonate content in sediments: reproducibility and comparability of results. J Paleolimnol
25:101-110
Kruse F, Boardman J, Huntington J (1999) Fifteen years of hyperspectral data: northern
Grapevine Mountains, Nevada. Proceedings of the 8th JPL Airborne Earth Science
Workshop: Jet Propulsion Laboratory Publication, JPL Publication, pp 99-17
Kruse F, Lefkoff A, Boardman J, Heidebrecht K, Shapiro A, Barloon P, Goetz A (1993) The
spectral image processing system (SIPS)—interactive visualization and analysis of imaging
spectrometer data. Remote Sens Environ 44:145-163
Menounos B (1997) The water content of lake sediments and its relationship to other physical
parameters: an alpine case study. Holocene 7:207-212
Rein B, Sirocko F (2002) In-situ reflectance spectroscopy–analysing techniques for high-
resolution pigment logging in sediment cores. Int J Earth Sci 91:950-954
Richards JA, Jia X (1999) Remote sensing digital image analysis: An introduction. Springer,
Berlin
Tukey JW (1977) Exploratory data analysis. Addison-Wesley Publishing Company, Reading,
MA, 688 pp
Zolitschka B (1998) A 14,000 year sediment yield record from western Germany based on
annually laminated lake sediments. Geomorphology 22:1-1712
top related