«Year: 1998 Endmember selection procedures for partial spectral unmixing of DAIS 7915 imaging spectrometer data in highly vegetated Areas Kneubühler, ...»
Zurich Open Repository and
University of Zurich
Endmember selection procedures for partial spectral unmixing of DAIS 7915
imaging spectrometer data in highly vegetated Areas
Kneubühler, Mathias; Schaepman, Michael E; Schläpfer, Daniel; Itten, Klaus I
Abstract: An intensively used agricultural test site in Switzerland is covered by the DAIS 7915 imaging
spectrometer in summer 1997. Three different methods of collecting endmembers for spectral unmixing are selected and compared against each other. The methods include a soil-vegetation-atmosphere-transfer approach (SVAT) based on a leaf optical properties model (PROSPECT) and a canopy model (SAIL), image based endmember selection and in-situ reflectance measurements using a ground spectroradiometer.
The presented methods are discussed and verified with an extensive ground truth. A rejection procedure for classification of unmixing results is proposed on the acceptance of constraint spectral unmixing results using the uncertainty, expressed by the RMS, of the endmember selection.
Posted at the Zurich Open Repository and Archive, University of Zurich ZORA URL: http://doi.org/10.5167/uzh-98454 Published Version
Originally published at:
Kneubühler, Mathias; Schaepman, Michael E; Schläpfer, Daniel; Itten, Klaus I (1998). Endmember selection procedures for partial spectral unmixing of DAIS 7915 imaging spectrometer data in highly vegetated Areas. In: 1st International EARSeL Workshop on Imaging Spectroscopy, Zurich, Switzerland, 6 October 1998 - 8 October 1998, 255-261.
ENDMEMBER SELECTION PROCEDURES PARTIAL
UNMIXING OF DAIS 7915 IMAGING SPECTROMETER DATA IN
HIGHLY VEGETATED AREASMathias Kneubuehler, Michael Schaepman, Daniel Schlaepfer and Klaus Itten Remote Sensing Laboratories Dept. of Geography University of Zurich Winterthurerstrasse 190 CH-8057 Zurich
ABSTRACTAn intensively used agricultural test site in Switzerland is covered by the DAIS 7915 imaging spectrometer in summer 1997. Three different methods of collecting endmembers for spectral unmixing are selected and compared against each other. The methods include a soil-vegetation-atmosphere-transfer approach (SVAT) based on a leaf optical properties model (PROSPECT) and a canopy model (SAIL), image based endmember selection and in-situ reﬂectance measurements using a ground spectroradiometer. The presented methods are discussed and veriﬁed with an extensive ground truth. A rejection procedure for classiﬁcation of unmixing results is proposed on the acceptance of constraint spectral unmixing results using the uncertainty, expressed by the RMS, of the endmember selection.
KEY WORDS: DAIS 7915, Spectral Unmixing, Hyperspectral Data, SVAT
1 INTRODUCTIONIn August 1997 the Digital Airborne Imaging Spectrometer (DAIS 7915) operated by DLR (German Aerospace Research Establishment) has been ﬂown over an intensively cultivated agricultural area, the Limpach Valley (470 m a.s.l.) located in Western Switzerland. The area covered by the DAIS is 2.5 x 10 km and consists mainly of crops, meadows and sugar beet.
The DAIS-7915 is a 79-channel high-resolution optical spectrometer that covers the wavelength-range between 0.5 µm and 12 µm using a Kennedy type scanning mechanism. The ﬁrst 72 channels cover the reﬂective part of the electromagnetic spectrum whereas the channels 73-79 are located in the MIR and TIR range. The DAIS is operated aboard DLR’s Dornier DO228 aircraft and has a swath angle of 52°, subdivided into 512 pixels per scanline . The ﬂight altitude was approximately 3840 m a.s.l, resulting in a ground sampling distance of 5.5 m in line direction. The ﬂightline is recorded using a differential GPS on board of the DO228 and on preselected ground control points .
Numerous spectroradiometric measurements on selected reference targets have been taken in the test area using a GER-3700 spectroradiometer. This 704 channel spectroradiometer is located in the 0.4 µm - 2.5 µm wavelength range. The GER-3700 is calibrated in the laboratory by RSL .
Mapping of the land cover and determination of the ﬁeld borders is based on aerial photography. This allows for assessing the quality of the spectral unmixing results. More than 90 ﬁelds with their land cover can therefore be identiﬁed on the DAIS image.
2 METHODOLOGYThe DAIS 7915 data are provided as radiance-calibrated data from DLR Oberpfaffenhofen.
The preprocessing of the imaging spectrometer data includes an MNF (Minimum Noise Fraction) transformation. The conversion to apparent reﬂectances is performed using an atmospheric correction program (ATCOR-2) . Since no in-situ meteorological data is available for the time of the datatake, the horizontal visibility is determined using a comparison of different modelling approaches with in-situ reference spectra. This procedure is carried out iteratively, based on IFCALI . The correction procedure strongly overestimates the water vapor due to laboratory calibration errors of the DAIS imaging data resulting in an overcorrection in the 940 nm water vapor absorption band .
The spectral library concept in general is based on the spectral variability of the selected endmembers. The dimensionality of the endmembers is determined in a four step approach: The measurement noise of the data is determined using MNF transforms. Non-linear effects like multiple scattering and transmittance through optically thin endmember layers of vegetation are not modeled and therefore act as unmodeled bias terms. To avoid a large number of endmembers in small proportions, the target area for spectral unmixing is reduced by assuming four major types of linear contributing endmembers such as the dry biomass (ripe wheat), soil, shade and vegetation (grass and sugar beet). Since large variations in the endmember spectra can contribute to low spectral contrast and limit the useful dimensionality of the data, endmembers should result from homogeneous data sources such as large homogeneous ﬁelds, especially if they are selected from imaging data .
Because of noisy SWIR bands of the DAIS and the importance of speciﬁc vegetation related absorption bands  (e.g. chlorophyll a and b, water absorption) and the position of the red-edge, the analysis focuses on the ﬁrst 40 channels of the DAIS, covering the 0.5 µm
- 1.8 µm. Major attention is given to agricultural land, so that the abovementioned four types of linear contributing endmembers are assumed to be present in the area under investigation in a statistically signiﬁcant amount. The three different approaches for the endmember collection
• image based selection using the inner part of a ﬁeld
• spectroradiometric measurements of reference targets
• SVAT modelling of endmember spectra Image based selection: Endmembers are selected within the inner part of a ﬁeld to avoid the selection of adjacency inﬂuenced border pixels. Shadow, that has a signal close to 0, because atmospherically corrected ground reﬂectance data contain no path scattered radiance, is introduced. Adjacency and directional effects in the shadow are assumed to be close to 0, too.
Spectroradiometric measurements of reference targets: Measurements of reference ﬁelds (e.g.
grass, sugar beet and ripe wheat) are convolved to the required wavelength of the DAIS 7915 channels.
Modelling endmember spectra: Endmembers of green vegetation (grass and sugar beet) are modeled using a SVAT approach. In order to simulate canopy reﬂectance, a single leaf reﬂectance model (PROSPECT ) and a canopy reﬂectance model (SAIL ) are combined.
PROSPECT uses as input a leaf mesophyll structure parameter N, chlorophyll content, leaf dry mass and leaf water content. It simulates leaf reﬂectance and leaf transmittance between 0.4 µm and 2.5 µm as a function of the abovementioned leaf optical properties. The output of the PROSPECT model is then used as input to the SAIL model. The SAIL model uses as input canopy variables (e.g. leaf area index, leaf angle distribution, leaf reﬂectance and transmittance), soil reﬂectance, the ratio between diffuse and direct irradiance and solar/view geometry (solar zenith angle, zenith view angle and relative angle between sun and view azimuth .
The PROSPECT parameters are calculated by inverting an averaged measured high resolution spectrum of vegetation (grass and sugar beet) and by minimizing the RMS-error between modeled and measured spectra. The SAIL model calculates canopy reﬂectance under different view zenith angles which become important when performing comparisons with sensors having a large scan angle such as the DAIS. Canopy reﬂectance depends on the sensor’s changing view zenith angle and the changing relative azimuth angle between sun and viewing direction. Since the main area of interest is completely ﬂat and lies within 13° to each side from nadir view, only nadir-modeled spectra are used for the modelling approach. However, Fig. 1 and Fig. 2 show the variation in canopy reﬂectance of grass and sugar beet resulting from changing view zenith angles. Since soil and dry biomass (ripe wheat) do not represent vital vegetation, these endmembers are not modeled, but substituted by the ﬁeld reﬂectance measurements. Fig. 1, Fig. 2 and Fig. 3 show the endmember spectra for grass, sugar beet, wheat and soil.
The evaluation of the different results of these three approaches of endmember selection for spectral unmixing are discussed focussing on well deﬁned veriﬁcation areas of sugar beet. Sugar beet is one of the predominant land cover in the observed area. Abundance maps of the sugar beet endmember are generated in a two step approach: First, based on the resulting RMS (root mean square error) of the spectral unmixing procedure, which is calculated for each pixel, areas having an RMS larger than (RMSmean+2 δ ( RMS ) ) were excluded from further interpretation. Large RMS values indicate poor unmixing results based on the selected endmembers. Second, endmember abundances lower than zero are also excluded.
Since in-ﬁeld variation can decrease the purity of an endmember group, an iterative application of this procedure can be used to obtain purer endmember spectra, especially in the case where they are derived from regions of interest of the imaging data. Fig. 4 clearly shows the inhomogenities of the sugar beet ﬁeld used for the collection of the sugar beet endmember.
FIGURE 4: In-ﬁeld variation of the sugar beet endmember ﬁeld as seen by the simultaneously ﬂown WAAC (Wide Angle Airborne Camera) The abundance maps of sugar beet are geocoded using a parametric geocoding approach (PARGE) including a DEM (Digital Elevation Model) and the attitude data of the aircraft .
3 RESULTS AND CONCLUSIONSSpectral mixture analysis is based on the assumption that most of the spectral variation within imaging data is caused by mixtures of a limited number of surface materials. Linear spectral unmixing assumes no interaction between materials and therefore models the observed spectral reﬂectance as linear combinations of pure endmembers . In this work, a weighted constrained linear unmixing is applied, allowing abundances to have values lower than 0 or greater than 1, but summing up to unity . As pointed out by Schanzer , abundances can take on values greater than 1 or lower than 0. If the response of a pixel is purer than the selected endmembers, endmember abundances greater than 1 and lower than 0 will occur.
Fig. 4 shows a geometrically corrected subset of the three unmixing results. The results from image based endmember selection (left), measured endmember spectra (middle) and modeled endmembers (right) are compared to each other focussing on sugar beet. Differences of abundances of the sugar beet endmember on the same ﬁelds are obvious for the three approaches.
FIGURE 5: Abundance maps of sugar beet from image derived (1), in-situ measured (2) and modeled (3) endmember spectra Image derived endmembers tend to be less pure than measured or modeled endmembers, since they are composed of a mixture of the observed vegetation as well as soil and shade present in the target area. Measuring endmember spectra focussing on the vegetated part itself or modelling them from biochemical, biophysical and geometrical parameters leads to purer spectra of the desired endmember. This fact can be seen in the three images very well. The image based approach (1) shows many ﬁelds with some amount of sugar beet abundance whereas no or less abundance of this endmember can be found in the same ﬁelds in the measured (2) and modeled (3) endmember approaches. Dark ﬁelds are sugar beet plantations. They show highest abundances of sugar beet for all three approaches, although the abundance values differ signiﬁcantly. Pixels as pure as the measured and modeled endmembers are hardly any found in the image.
The scene based approach instead bases on impure endmembers. The ﬁeld where the endmembers for sugar beet are collected (see Fig. 4) is obviously a mixture of various contributing species like soil, shade, weed and sugar beet. Therefore, evaluation ﬁelds with a higher amount of sugar beet lead to a purer response than the endmember itself. This is the case of feasible endmember collection . Abundance values greater than 1 occur. This forces parts of the remaining endmembers to become negative.
To solve this problem, an iterative approach of scene based endmember collection is applied to the data. A ﬁrst unmixing result allows identiﬁcation of purer endmembers than the initial ones. Based on the resulting RMS of this ﬁrst spectral unmixing procedure, pixels of the veriﬁcation area that have an RMS larger than (RMSmean+2 δ ( RMS ) ) are excluded.