SPECTRAL UNMIXING FOR INFORMATION EXTRACTION R.Kancheva, D.Borisova* STIL-BAS Acad.G.Bonchev str., bl.3, 1113 Sofia, Bulgaria E-mail:
[email protected] /
[email protected] KEY WORDS: Remote Sensing, Agriculture Crop, Data Analysis, Theory Application, Land Cover Classification
ABSTRACT: “From pixels to processes” – this the shortest and most essential formulation of all studies, experiments and investigations carried out in the field of Earth remote sensing observations. This formulation reveals two main directions of data interpretation: the first related to classification and feature retrieval, and the second associated with multi-temporal aspects of remotely sensed data and concerning change detection and processes tracking and modeling. State assessment, trends forecasting and predictions are the goals of the environmental and land cover monitoring. Further implementation of the investigation results resemble the decision making and problem-solving nature of remotely sensed data outputs. The development of efficient technologies for data analysis is one of the most challenging issues that the remote sensing community is facing. Matters of data reduction, processing algorithms accuracy, information amount, cost and time saving determine the efficiency of data analysis. The importance of this issue is directly connected with the ever-increasing quantity of data provided by numerous optical, thermal and microwave sensors, with their synergistic use as well as with the accuracy of data processing algorithms and results verification. With all this in mind we present here some results from a study of different spectral unmixing techniques over rock-soil-vegetation objects in relation to mixtures decomposition, objects type and proportions determination and biophysical properties retrieval. Experimental data from field and laboratory spectral reflectance measurements in the visible and near infrared band have been used, various decomposition methods (linear unmixing, clustering, colorimetric analyses, etc.) have been applied and evaluated, comparison between empirical and simulation models has been performed.
1. INTRODUCTION The development of efficient technologies for data analysis is one of the most challenging issues that the remote sensing community is facing. Matters of data reduction, processing algorithms accuracy, information amount, cost and time saving determines the efficiency of data analysis. The importance of this issue is directly connected with the ever-increasing quantity of data provided by numerous airborne, field and laboratory operated sensors, with their synergistic use as well as with the accuracy of data processing algorithms and results verification. With all this in mind we present here some results from a study of different spectral unmixing techniques over rock-soilvegetation objects in relation to mixtures decomposition, objects type and proportions determination and biophysical properties retrieval. Experimental data from field and laboratory spectral reflectance measurements in the visible and near infrared band have been used, various decomposition methods (linear unmixing, colorimetric analyses, clustering, etc.) have been applied and evaluated, comparison between empirical and simulation models has been performed. Spectral linear unmixing is efficient approach to the spectral decomposition of multichannel remotely sensed data. A main problem to its process is that the number of spectral components (end-members) has to be correctly distinguished. The applied methods are studied for their potential to decompose mixture components from spectral reflectance data in two cases.
* Corresponding author.
In the first case on the example of soil-vegetation covers the influence of the different soil background and plant senescence (changes in chlorophyll content during the growing period) on vegetation reflectance and color features are analyzed. Spectral transformation techniques and colorimetrical analysis of experimental and modeled data sets are used. Some results of this study concerning the estimation of green vegetation fraction cover are shown. In the second case on the example of rock-soil-vegetation land cover the relationships between spectral reflectance of mixed pixel and fraction cover of rock, soil and vegetation therein and the evaluation the possibility of using spectral mixture decomposition in relation to their type and proportion determination are described. 2. MATERIALS AND METHODS
Study case one Ground-based field and greenhouse reflectance measurements of wheat, barley, alfalfa, grassland with different soil background and degree of senescence are carried out with multichannel radiometers (Krumov, 1983; Iliev, 2000) in the VIS and NIR spectral bands. The soil variety was presented by dark and light soils (chernozem, brown, grey forest, alluvial). The spectral reflectance curves of some of them are presented in Fig.1a illustrating the wide range of soil reflectance signatures. The reflectance spectra of meadow with different vegetation fraction cover are shown in Fig.1b where the impact of the soil type: dark chernozem (1) and light alluvial soil (2) is
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. 34, Part XXX observed. Some spectral transformations well-known as vegetation indices (Table 1) used in vegetation studies are calculated in the green (G – 550 nm), red (R – 670 nm) and near infrared (NIR – 800 nm) bands.
In the case of soil and green vegetation, for instance,
N
VI
N
VI
1 2
(NIR-R)/(NIR+R) NIR/R
3 4
(G-R)/(G+R) G+NIR-2R
rsv = pv rv + (1 − pv )rs
(2)
rsv = p v (rv − rs ) + rs
(3)
Table 1. Vegetation indices
∑p
i
considering that
=1
i
:
30
The tristimulus values X,Y,Z, chromaticity coefficients x,y,z and dominant wavelength λd of the measured objects and modeled mixtures are computed in the spectral range 450-750 nm according to the CIE 1964 methods and D65 light source.
reflectance, %
alluvial 20
grey
According to the unmixing theory the following equations are worked out (the same being true for Y, Z and W=X+Y+Z) (Mishev, 1992; Kancheva, 2003):
10
chernozem 0
500
600
700
wavelength, nm
800
X sv = ∑ D65 [ pv (rv − rs ) + rs ] xΔλ
(4)
X sv = pv ( X v − X s ) + X s
(5)
pv (X v − X s ) + X s pv (Wv − Ws ) + Ws
(6)
λ
a) 40
2b 1b Pv=0.8
reflectance, %
30
2a Pv=0.2 1a
20
xsv =
10
0
500
600
700
wavelength, nm
The (xsv,ysv) coordinates (Eq. (6)) define the position of soilvegetation mixtures on the color locus and depend on the fraction cover of the end-members pv. For a three-component mixture including plant dried foliage:
800
b) Figure 1. Reflectance spectra of different soils (a) and soilgreen vegetation mixtures (b) The variety of green and dried vegetation fraction covers the latter related to chlorophyll decrease in plants was achieved during plant development stages or modeled from bare soil and vegetation full-canopy reflectance using the unmixing theory (Mishev, 1991):
rΣ (λ ) =
where
∑ p r (λ ) i i
i
(1)
rΣ (λ ) are the resulting spectral reflectance signatures of the mixture,
ri (λ )
- the reflectance of the components (endmembers) composing the mixture,
pi
- components’ relative amounts (fraction cover).
xsvd =
pv ( X v − X s ) + p d ( X d − X s ) + X s pv (Wv − Ws ) + pd (Wd − Ws ) + Ws
xsvd =
pv ( X v − X d ) + ps ( X s − X d ) + X d pv (Wv − Wd ) + ps (Ws − Wd ) + Wd
(7)
(8)
Statistical data processing was carried out to investigate variances in vegetation reflectance and color features due to different soil background and plant cover (green and dried fractions). Correlation and regression analysis was applied to the experimental and modeled data sets to reveal the relationships between 1) the vegetation indices and green vegetation cover pv, and 2) the dominant wavelength λd and green vegetation cover pv.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. 34, Part XXX
Study case two Region of interest (ROI) in the second study area AssarelMedet is selected based on TM dataset. Assarel-Medet – an open pit copper mine (Figure 2) is situated on an area of 20 thousand decares, about 1000 m above sea level, 90 km southeast from the capital of Bulgaria, Sofia, in the Sashtinska Sredna Gora Mountain. During field campaign 2005 granite, brown soil and grass samples are collected. Mineral and chemical reference analyses of rock and soil samples are carried out. Field and laboratory spectral measurements are also implemented. Pinkish mediumgrained granite samples consist of average 50% orthoclase, 35% quartz, 15% plagioclase, 1% biotite and 1% magnetite. The parent material of the brown sandy loam soil sample is granite. The soil sample has moderate to coarse prismatic structure, very hard and friable with neutral pH (6.5) 1% organic carbon, 12% clay, 25% silt and 62% sand. The grassland vegetation type is presented by Green Rye grass.
ues from 0.86 to 0.97. As the normalized difference of the measured spectral reflectance in the red and near infrared band is traditionally and most widely used, the presented below examples refer mainly to this spectral index. In Figure 3 the statistical relationships of NDVI = (NIR-R)/(NIR+R) and barley canopy cover (fraction cover) are shown. The dependences are derived separately for the grey (1) and chernozem (2) soil plots. Regression analysis was applied to reveal the relationship between λd and pv. Strong correlation was found between the green canopy cover (fraction cover) and the dominant wavelength of soil-vegetation mixtures. The fitted dependences pv=f(λd) are second degree polynomials with coefficients of determination 0.94 and 0.91 (at p