Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

A NOVEL APPROACH FOR ABUNDANCE ESTIMATION USING DISCONTINUITY

PRESERVING PRIOR

Jignesh R. Patel, Manjunath V. Joshi Jignesh S. Bhatt

Dhirubhai Ambani Institute of Indian Institute of


Information and Communication Technology, Information Technology,
Gandhinagar, India. Vadodara, India.

ABSTRACT In LMM, mixed data can be modeled for single input


vector of reflectance as,
In this paper, we propose an algorithmic approach to esti- r = M α + η, (1)
mate fractions (abundances) of materials (endmembers) in a where, r ∈ Rl×1 denotes the spectral measurement, M ∈
pixel by considering linear mixture model (LMM) and prior Rl×p is the endmember matrix, α ∈ Rp×1 is called the
knowledge of endmembers. We propose the use of inho- abundance vector. Here, l represents the number of spectral
mogeneous Gaussian Markov random field (IGMRF) that bands, p is the number of endmembers and η corresponds
captures the smoothness as well as preserves the discon- to independent and identically distributed (IID) white Gaus-
tinuities among abundance values. We obtain the IGMRF sian noise. We know that the α estimation requires sum-to-
parameters for the proposed prior using the initial esti- one and no-negativity constraints. To this end, we use a
mate of abundances and both the abundances and IGMRF similar approach as given in FCLS [4] for the sum-to-one
parameters are refined by optimizing an objective func- constraint and the non-negativity constraint is taken care
tion. A two-step iterative approach is proposed to obtain while optimization of the objective function by considering
the final estimates of both the abundance maps and their only the positive values while updating the abundances.
prior parameters. In order to demonstrate the efficiency In general, abundance maps are spatially smooth, which
of the proposed approach, we conduct experiments on the in practice is captured by Markov random field (MRF)
synthetic hyperspectral images (HSIs) with different noise based prior. However, the use of homogeneous MRF does
levels as well as on the real HSIs and compare the result not take into account the presence of discontinuities in the
with other state-of-the-art approaches. abundance values. To this end, IGMRF prior can be used to
Index Terms— Abundance estimation, Spectral un- capture the smoothness and also preserve the discontinuity
mixing, Fully constrained least squares (FCLS), Inhomo- [5]. However, the estimation of these parameters requires
geneous Gaussian Markov random field (IGMRF). the availability of true abundances which themselves have
to be estimated. In our work, we estimate them by us-
1. INTRODUCTION ing an initial abundance map and modify them iteratively
Hyperspectral images (HSIs) represent the collection of re- while refining the abundances. We used matched filter
flectance in hundreds of contiguous electromagnetic bands (MF) [6] based approach for finding the initial estimate of
of narrow bandwidth. The rich spectral details of HSI help abundances maps which are then used in computing the
us in solving numerous problems such as classification of parameters for the prior. To get the refined final estimate of
the scene data, target detection, anomaly detections, etc. abundances an objective function consisting of data fitting
The use of high spectral resolution image results in very and prior terms is formulated and is optimized using Adam
high discrimination capabilities. However, due to the con- optimization [7]. Using an iterative approach, refined abun-
straints such as hardware, recording time, and the signal dance maps are used to reinitialize the IGMRF parameters
to noise ratio (SNR), the HSIs have low spatial resolution and then again the optimization is carried out in order to
making the sensor reflectance to represent mixture of more obtain final abundances. Doing so, as a by product, we also
than one material. The process of decomposition of mixed get the refined IGMRF parameters leading to their better
pixel (reflectance) into endmembers and their abundances estimates.
is called spectral unmixing. The concept of unmixing re- 2. DETAILED DESCRIPTION OF THE PROPOSED
quires three steps: (a) estimation of the number of endmem- APPROACH
bers, (b) extraction of endmembers values, and (c) estima- The block diagram of the proposed approach is shown in
tion of their abundances [1]. In this work, we estimate the Fig. 1. We formulate this estimation problem in an energy
abundance maps and assume that both the number of end- minimization framework. Given the hyperspectral data and
members and their values (signatures) are available. Virtual the endmembers, we estimate the abundances α by mini-
dimensionality (VD) [2] is a standard algorithm to identify mizing the following energy function,
the number of endmembers, while Vertex component anal- E(α) = ED (α) + EP (α), (2)
ysis (VCA) [3] extracts endmembers of the data by consid- where, ED (α) represents the data fitting term measuring
ering convexity of the data. the reconstruction error between the given and the recon-
1
HSI Endmembers
0.8
Initial abundance estimating

bxIGM RF \bx .
using MF 0.6

0.4
IGMRF bxIGM RF
parameter initialization 0.2

bx
Adam optimization 0.2 0.4 0.6 0.8 1
dx

Estimated abundance maps Fig. 2: Variation of IGMRF parameter bxIGM RF and the
proposed parameter bx with the difference dx between two
Fig. 1: Block diagram of the proposed approach. neighboring abundances.
structed reflectance. EP (α) represents the prior term that
2.2. Estimation of final abundance maps
constraints the smoothness and discontinuity of abundance
As already stated earlier, we estimate the abundance maps
values in the spatial domain. Using the LMM given in
by using energy minimization framework which requires
Equation (1), ED (α) can be easily obtained as,
X data and the prior terms. Using these terms given in Equa-
ED (α) = kr(x, y) − M α(x, y)k2 , (3) tion (3) and Equation (4), the final energy function to be
(x,y) minimized is Xgiven by,
where, r(x, y) represents the spectral measurement at (x, y) E(α) = kr(x, y) − M α(x, y)k2 +
pixel location and α(x, y) represents the abundance vector (x,y)

consisting of abundance values for each endmember. We


X X
bx
(x,y,p) (α(x − 1, y, p) − α(x, y, p))
2 (7)
define the prior term EP (α) using the modified IGMRF as (x,y) p
follows. + by(x,y,p) (α(x, y − 1, p) − α(x, y, p))2 .
2.1. Modified IGMRF prior
The fact that the abundance maps have both smoothness The function given in Equation (7) is convex and hence can
as well as discontinuities motivates us to use IGMRF as a be easily minimized by using gradient-descent optimization
prior. While using IGMRF prior, one can write the prior method. To carry out the optimization, We use Adam opti-
term EP (α) as the sum of squares of finite difference ap- mizer [7] which computes the adaptive learning rates from
proximations to the first order derivatives of abundances at the estimates of first and second moments of the gradients
each pixel location (x, y) for each of the endmember p [5] of energy function w.r.t abundances. Once the abundances
as follows, X X are estimated, we iterate the following two steps until con-
EP (α) = bx
(x,y,p) (α(x − 1, y, p) − α(x, y, p))
2
vergence in order to get the final abundance maps
(x,y) p (4) Step 1: Compute bx and by parameters using Equation
+by(x,y,p) (α(x, y − 1, p) − α(x, y, p))2 , (6) by keeping α fixed.
where, bx and by are the spatially adaptive IGMRF pa- Step 2: Optimize Equation (7) for α by keeping the bx
rameters in horizontal and vertical directions, respectively. and by fixed.
In [5], the authors used a maximum likelihood estimation 3. EXPERIMENTAL RESULTS
method to derive the IGMRF parameters as, We first conduct the experiment on synthetic HSIs gener-
xIGM RF
1
b(x,y,p) = , ated using three spectral signatures of the U.S. Geological
max(4(αin (x − 1, y, p) − αin (x, y, p)))2 , 4)
1
(5) Survey (USGS) digital spectral library [8]. We next show
yIGM RF
b(x,y,p) = , the experiments on the two real data sets consisting of
max(4(αin (x, y − 1, p) − αin (x, y, p)))2 , 4)
Urban data captured by Hyper-spectral Digital Imagery
where, αin represents approximation to α which corre-
Collection Experiment (HYDIC) [9] and Cuprite data col-
sponds to initial estimate in our work. Since the values of
lected by the Airborne Visible/Infrared Imaging Spectrom-
abundances vary between 0 and 1, the IGMRF parameters
eter (AVIRIS) [10] used in the hyperspectral unmixing
computed using Equation (5) also vary in the same range,
[11]. The IGMRF parameters were initialized with those
thus imposing the same penalty for differences, which is
computed using the abundances estimated from the MF
illustrated in Fig. 2. Hence, the computed parameters us-
based approach, and with this our algorithm converges in
ing Equation (5) fail to preserve the discontinuity in the
four iterations. We compare our proposed approach with
abundances. We modify the computation of parameters as,
1 other state-of-the-art approaches. In order to check how our
bx(x,y,p) = 1 − , method performs on any initial estimate, we also performed
1 + e−5×|(αin (x−1,y,p)−αin (x,y,p))|
(6) the experiments by initializing the IGMRF parameters ob-
1
by(x,y,p) = 1 − . tained using the initial estimate of abundances obtained by
1 + e−5×|(αin (x,y−1,p)−αin (x,y,p))| FCLS [4] and Maxent [12] approaches. In this case, our
In Fig. 2, we also show the variation of bx (modified param- algorithm converged in less than four iterations. The per-
eter) w.r.t variation in dx = |αin (x−1, y, p)−αin (x, y, p)|. formance comparison using quantitative method was done
As seen in Fig. 2, the proposed parameter now has a larger by using the root mean squared error (RMSE) [13] and
penalty for smaller differences (smooth region) and smaller the spectral angle mapper (SAM) [14] as the quantitative
penalty for higher differences (edges).
Ammonioalunite Brucite Andradite Table 1: Quantitative comparison on synthetic data
Initial SNR
Measure Algorithm
estimate -5dB 0dB 5dB 10dB 20dB
Maxent [12] - 0.0900 0.0562 0.0368 0.0245 0.0129
FCLS [4] - 0.1002 0.0600 0.0344 0.0194 0.0068
DGAE [18] - 0.3383 0.2473 0.1198 0.0370 0.0072
RMSE [13]
MF 0.0971 0.0528 0.0284 0.0160 0.0062
Proposed Maxent 0.0874 0.0498 0.0280 0.0163 0.0065
(a) Ground truth FCLS 0.0907 0.0508 0.0277 0.0157 0.0063
Maxent [12] - 0.1678 0.1044 0.0681 0.0443 0.0231
FCLS [4] - 0.1890 0.1125 0.0652 0.0370 0.0127
DGAE [18] - 0.5431 0.4185 0.2236 0.0647 0.0134
SAM [14]
MF 0.1810 0.0981 0.0539 0.0301 0.0113
Proposed Maxent 0.1626 0.0924 0.0530 0.0308 0.0117
FCLS 0.1698 0.0947 0.0523 0.0298 0.0115

Table 2: Quantitative comparison on real data


(b) FCLS [4]
Measure Algorithm Initial estimate Urban Data [9] Curprite Data [10]
Maxent [12] - 0.0687 0.0375
FCLS [4] - 0.0682 0.0369
cSUNSAL [16] - 0.0677 0.0438
SPU [17] - 0.0682 0.0369
DRE
DGAE [18] - 0.0699 0.0374
MF 0.0411 0.0211
Proposed Maxent 0.0522 0.0209
(c) Proposed FCLS 0.0402 0.0209
Fig. 3: Abundance maps for synthetically generated data
for 5 dB SNR. the noisy bands representing the water adsorptions were re-
moved and the experiment was performed on 162 bands
measures. The expression for RMSE and SAM are given only. We also conducted experiment on 188 bands of HSI
by the following
v
equations, data acquired by AVIRIS sensor [10] and for this exper-
u
u1 X p
αT α iment, we used sub-images of the size 250 × 190 pixels
ci )2 , SAM = cos−1
(αi − α
b
RM SE = t . (8) having 224 bands with 12 endmembers after removing the
p i=1 kαkkb
αk
Here, αi and α̂i represent the abundance values of true and noisy bands. Since the true abundances are not available
estimated abundances where as α and α̂ correspond to the here, we used the data reconstruction error (DRE) mea-
respective vectors. sured in terms of RMSE at each pixel between the original
and the reconstructed HSI as the quantitative measure. As
3.1. Experiments on synthetic data
given in Table 2, the results for the proposed approach are
We generate the synthetic HSIs by using the linear mixing better than other state-of-the-art methods. Note that, in the
of spectral signature recorded using 224 bands of the USGS case of synthetic data, where the noise is also simulated,
library [15]. Here, the size of each synthetic HSI is 75 × 75 our approach with MF based initialization is slightly infe-
pixels which consist of three materials: namely, Ammo- rior when compared to Maxent at very low SNR (-5dB).
nioalunite NMNH145596, Brucite HS247.3B, and Andra- However, for real data, our approach performed better even
dite WS487. The experiments are performed at different when initialized with MF method. The estimated abun-
SNR levels, and the quantitative comparison is shown in dance maps for Urban data are shown in the Fig. 4, where
Table 1. Here, in addition to Maxent [12] and FCLS [4], the we observe that the abundance maps estimated using the
recently proposed methods cSUNSAl [16], SPU [17], and proposed approach are closer to the ground truth when
DGAE [18] are also considered for comparison. However, compared to the FCLS approach. Before we conclude, we
the performance of the methods cSUNSAl [16] and SPU would like to mention few points regarding computational
[17] is similar to FCLS [4] and hence we have not included complexity. Though, the proposed approach is iterative,
in these in the Table 1. The use of MF based approach for the computation time is not significantly high since our
initial abundance estimation performs better than the others algorithm converged in very few iterations. Note that, the
except for SNR = -5dB where Maxent [12] performs better. other state-of-the-art approaches [4, 12, 16, 17, 18] are
However in this case, our approach with Maxent initializa- not iterative and hence are slightly better in terms of time
tion performs better when compared to Maxent itself. The complexity. However, our method considers the spatial
results also show that initialization with Maxent [12] and correlation of abundances at every location in the form of
FCLS [4] perform better. The estimated abundance maps IGMRF parameters which is not done in other approaches.
using the proposed as well as the FCLS methods are shown
in Fig. 3 for 5 dB SNR. Looking at Fig. 3, we observe 4. CONCLUSION
that the homogeneous region in abundance maps estimated In this paper, we have proposed a novel approach to esti-
using the proposed approach appear smoother when com- mate abundance maps by energy minimization framework.
pared to the FCLS and also our method better preserves the A new discontinuity preserving prior is proposed by mod-
discontinuities among the abundances as evident at the bor- ifying the expression for IGMRF parameters. A two-step
ders of the abundance squares. iterative algorithm is used to refine the abundance maps
3.2. Experiments on real data as well as the parameters. The experiments conducted on
We applied our algorithm on Urban hyperspectral images synthetic and the real HSIs show that our method performs
of HYDIC [9] of size 307 × 307 pixels having 210 bands better when compared to other state-of-the-art approaches
with 4 endmembers (Asphalt, Grass, Tree, and Roof). Here, even at very low SNR.
Asphalt Grass Tree Roof [8] Roger N Clark, Gregg A Swayze, Richard Wise,
K Eric Livo, T Hoefen, Raymond F Kokaly, and
Stephen J Sutley, “Usgs digital spectral library
splib06a,” US geological survey, digital data series,
vol. 231, 2007.
(a) Ground truth [9] S. J. Ford, J. C. McGlone, S. D. Cochran, J. A. Shufelt,
W. A. Harvey, and D. M. McKeown, “Analysis of hy-
dice data for information fusion in cartographic fea-
ture extraction,” in Geoscience and Remote Sens-
ing Symposium Proceedings, 1998. IGARSS ’98. 1998
IEEE International, Jul 1998, vol. 5, pp. 2702–2706
(b) FCLS [4] vol.5.
[10] Gregg Vane, Robert O Green, Thomas G Chrien,
Harry T Enmark, Earl G Hansen, and Wallace M
Porter, “The airborne visible/infrared imaging spec-
trometer (aviris),” Remote sensing of environment,
(c) Proposed vol. 44, no. 2-3, pp. 127–143, 1993.
Fig. 4: Abundance maps for URBAN hyperspectral data of
[11] F. Zhu, Y. Wang, B. Fan, S. Xiang, G. Meng, and
[9].
C. Pan, “Spectral unmixing via data-guided sparsity,”
5. REFERENCES IEEE Transactions on Image Processing, vol. 23, no.
12, pp. 5412–5427, Dec 2014.
[1] J. M. Bioucas-Dias, A. Plaza, N. Dobigeon, M. Par-
ente, Q. Du, P. Gader, and J. Chanussot, “Hyper- [12] L. Miao, H. Qi, and H. Szu, “A maximum entropy ap-
spectral unmixing overview: Geometrical, statistical, proach to unsupervised mixed-pixel decomposition,”
and sparse regression-based approaches,” IEEE Jour- IEEE Transactions on Image Processing, vol. 16, no.
nal of Selected Topics in Applied Earth Observations 4, pp. 1008–1021, April 2007.
and Remote Sensing, vol. 5, no. 2, pp. 354–379, April [13] Z. Wang and A. C. Bovik, “Mean squared error: Love
2012. it or leave it? a new look at signal fidelity measures,”
[2] Chein-I Chang and Qian Du, “Estimation of num- IEEE Signal Processing Magazine, vol. 26, no. 1, pp.
ber of spectrally distinct signal sources in hyperspec- 98–117, Jan 2009.
tral imagery,” IEEE Transactions on Geoscience and [14] Roberta H Yuhas, Alexander FH Goetz, and Joe W
Remote Sensing, vol. 42, no. 3, pp. 608–619, March Boardman, “Discrimination among semi-arid land-
2004. scape endmembers using the spectral angle mapper
(sam) algorithm,” 1992.
[3] J. M. P. Nascimento and J. M. B. Dias, “Vertex com-
ponent analysis: a fast algorithm to unmix hyperspec- [15] RN Clark, GA Swayze, A Gallagher, T King, and
tral data,” IEEE Transactions on Geoscience and Re- W Calvin, “The us geological survey digital spec-
mote Sensing, vol. 43, no. 4, pp. 898–910, April 2005. tral library: Version 1: 0.2 to 3.0 µm. open file report
93-592,” US Geological Survey, 1993.
[4] D. C. Heinz and Chein-I-Chang, “Fully constrained
least squares linear spectral mixture analysis method [16] J. M. Bioucas-Dias and M. A. T. Figueiredo, “Alter-
for material quantification in hyperspectral imagery,” nating direction algorithms for constrained sparse re-
IEEE Transactions on Geoscience and Remote Sens- gression: Application to hyperspectral unmixing,” in
ing, vol. 39, no. 3, pp. 529–545, Mar 2001. 2010 2nd Workshop on Hyperspectral Image and Sig-
nal Processing: Evolution in Remote Sensing, June
[5] A. Jalobeanu, L. Blanc-Feraud, and J. Zerubia, “An 2010, pp. 1–4.
adaptive gaussian model for satellite image deblur-
ring,” IEEE Transactions on Image Processing, vol. [17] R. Heylen, D. Burazerovic, and P. Scheunders, “Fully
13, no. 4, pp. 613–621, April 2004. constrained least squares spectral unmixing by sim-
plex projection,” IEEE Transactions on Geoscience
[6] J. W. Boardman and F. A. Kruse, “Analysis of imaging and Remote Sensing, vol. 49, no. 11, pp. 4112–4122,
spectrometer data using n -dimensional geometry and Nov 2011.
a mixture-tuned matched filtering approach,” IEEE
Transactions on Geoscience and Remote Sensing, vol. [18] H. Pu, W. Xia, B. Wang, and G. M. Jiang, “A fully
49, no. 11, pp. 4138–4152, Nov 2011. constrained linear spectral unmixing algorithm based
on distance geometry,” IEEE Transactions on Geo-
[7] Diederik Kingma and Jimmy Ba, “Adam: A science and Remote Sensing, vol. 52, no. 2, pp. 1157–
method for stochastic optimization,” arXiv preprint 1176, Feb 2014.
arXiv:1412.6980, 2014.

You might also like